diff --git a/Cargo.lock b/Cargo.lock
index d18d8bebb..90c0664ee 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -11969,6 +11969,17 @@ dependencies = [
  "yaak-templates",
 ]
 
+[[package]]
+name = "yaak-query"
+version = "0.1.0"
+dependencies = [
+ "serde",
+ "serde-wasm-bindgen",
+ "serde_json",
+ "ts-rs",
+ "wasm-bindgen",
+]
+
 [[package]]
 name = "yaak-rpc"
 version = "0.0.0"
diff --git a/Cargo.toml b/Cargo.toml
index 40557a8ea..5e29d0da6 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -19,6 +19,7 @@ members = [
   "crates/yaak-mcp",
   "crates/yaak-models",
   "crates/yaak-plugins",
+  "crates/yaak-query",
   "crates/yaak-sse",
   "crates/yaak-sync",
   "crates/yaak-templates",
@@ -96,6 +97,7 @@ yaak-lifecycle = { path = "crates/yaak-lifecycle" }
 yaak-mcp = { path = "crates/yaak-mcp" }
 yaak-models = { path = "crates/yaak-models" }
 yaak-plugins = { path = "crates/yaak-plugins" }
+yaak-query = { path = "crates/yaak-query" }
 yaak-sse = { path = "crates/yaak-sse" }
 yaak-sync = { path = "crates/yaak-sync" }
 yaak-templates = { path = "crates/yaak-templates" }
diff --git a/apps/yaak-client/components/Sidebar.tsx b/apps/yaak-client/components/Sidebar.tsx
index 80d474656..cb3044736 100644
--- a/apps/yaak-client/components/Sidebar.tsx
+++ b/apps/yaak-client/components/Sidebar.tsx
@@ -63,10 +63,9 @@ import { isSidebarFocused } from "../lib/scopes";
 import { navigateToRequestOrFolderOrWorkspace } from "../lib/setWorkspaceSearchParams";
 import type { ContextMenuProps, DropdownItem } from "./core/Dropdown";
 import { ContextMenu, Dropdown } from "./core/Dropdown";
-import { type FieldDef, filter } from "@yaakapp-internal/ui/editor";
-import type { Ast } from "./core/Editor/filter/query";
-import { evaluate, parseQuery } from "./core/Editor/filter/query";
-import { formatFieldFilter } from "./core/Editor/filter/format";
+import type { Node as QueryNode, Schema } from "@yaakapp-internal/query";
+import { formatField, parseQuery, QueryMatcher } from "@yaakapp-internal/query";
+import { queryFilter } from "./core/Editor/filter/extension";
 import { HttpMethodTag } from "./core/HttpMethodTag";
 import {
   Icon,
@@ -111,7 +110,7 @@ function Sidebar({ className }: { className?: string }) {
   const activeWorkspaceId = useAtomValue(activeWorkspaceAtom)?.id;
   const treeId = `tree.${activeWorkspaceId ?? "unknown"}`;
   const filterText = useAtomValue(sidebarFilterAtom);
-  const [tree, allFields, emptyFilterSuggestions] = useAtomValue(sidebarTreeAtom) ?? [];
+  const [tree, fieldValues, emptyFilterSuggestions] = useAtomValue(sidebarTreeAtom) ?? [];
 
   const wrapperRef = useRef<HTMLElement>(null);
   const treeRef = useRef<TreeHandle>(null);
@@ -552,18 +551,18 @@ function Sidebar({ className }: { className?: string }) {
   const filterCompartmentMountExtRef = useRef<Extension | null>(null);
   if (filterCompartmentMountExtRef.current == null) {
     filterCompartmentMountExtRef.current = filterLanguageCompartmentRef.current.of(
-      filter({ fields: allFields ?? [] }),
+      queryFilter({ schema: sidebarFilterSchema, values: fieldValues }),
     );
   }
 
   useEffect(() => {
     const view = filterRef.current;
     if (!view) return;
-    const ext = filter({ fields: allFields ?? [] });
+    const ext = queryFilter({ schema: sidebarFilterSchema, values: fieldValues });
     view.dispatch({
       effects: filterLanguageCompartmentRef.current.reconfigure(ext),
     });
-  }, [allFields]);
+  }, [fieldValues]);
 
   if (tree == null || hidden) {
     return null;
@@ -851,16 +850,35 @@ function setSidebarFilterText(text: string) {
   jotaiStore.set(sidebarFilterAtom, { text, key: `${Math.random()}` });
 }
 
-function getSidebarSuggestionValue(ast: Ast | null) {
+/** The fields `getItemFields` fills in. Bare words match `getItemText` by word prefix */
+const sidebarFilterColumns = [
+  "method",
+  "url",
+  "name",
+  "type",
+  "folder",
+  "grpc_service",
+  "grpc_method",
+];
+const sidebarFilterSchema: Schema = {
+  fields: sidebarFilterColumns.map((name) => ({ name, kind: { type: "text" } })),
+  freeText: "fuzzy",
+};
+
+function getSidebarSuggestionValue(ast: QueryNode | null, query: string) {
   if (ast == null) return null;
 
   if (ast.type === "Term" || ast.type === "Phrase") {
-    const value = ast.value.trim();
+    const value = ast.text.trim();
     return value.length > 0 ? value : null;
   }
 
   if (ast.type === "Field") {
-    const value = ast.value.trim();
+    const raw =
+      ast.value.type === "text"
+        ? ast.value.text
+        : query.slice(ast.valueSpan.start, ast.valueSpan.end);
+    const value = raw.trim();
     return value.length > 0 ? value : null;
   }
 
@@ -882,7 +900,7 @@ const sidebarSuggestionFieldOrder = [
 ];
 
 const sidebarTreeAtom = atom<
-  [TreeNode<SidebarModel>, FieldDef[], SidebarFilterSuggestion[]] | null
+  [TreeNode<SidebarModel>, Record<string, string[]>, SidebarFilterSuggestion[]] | null
 >((get) => {
   const allModels = get(memoAllPotentialChildrenAtom);
   const activeWorkspace = get(activeWorkspaceAtom);
@@ -903,15 +921,18 @@ const sidebarTreeAtom = atom<
     return null;
   }
 
-  const queryAst = parseQuery(filter.text);
-  const suggestionValue = getSidebarSuggestionValue(queryAst);
+  const suggestionValue = getSidebarSuggestionValue(parseQuery(filter.text), filter.text);
+  const matcher = new QueryMatcher(filter.text, sidebarFilterSchema);
+
+  // Every request goes into one flat table (its text, then a cell per column), matched in a
+  // single call after the tree is built
+  const leaves: TreeNode<SidebarModel>[] = [];
+  const cells: (string | null)[] = [];
 
-  // returns true if this node OR any child matches the filter
   const allFields: Record<string, Set<string>> = {};
   const suggestionFields = new Set<string>();
-  const build = (node: TreeNode<SidebarModel>, depth: number): boolean => {
+  const build = (node: TreeNode<SidebarModel>, depth: number) => {
     const childItems = childrenMap[node.item.id] ?? [];
-    let matchesSelf = true;
     const fields = getItemFields(node);
     const model = node.item.model;
     const isLeafNode = !(model === "folder" || model === "workspace");
@@ -929,12 +950,12 @@ const sidebarTreeAtom = atom<
       }
     }
 
-    if (queryAst != null) {
-      matchesSelf = isLeafNode && evaluate(queryAst, { text: getItemText(node.item), fields });
+    if (isLeafNode && !matcher.isEmpty) {
+      leaves.push(node);
+      cells.push(getItemText(node.item));
+      for (const column of sidebarFilterColumns) cells.push(fields[column] ?? null);
     }
 
-    let matchesChild = false;
-
     // Recurse to children
     node.children = !isLeafNode ? [] : undefined;
 
@@ -952,19 +973,21 @@ const sidebarTreeAtom = atom<
 
       for (const item of childItems) {
         const childNode = { item, parent: node, depth };
-        const childMatches = build(childNode, depth + 1);
-        if (childMatches) {
-          matchesChild = true;
-        }
+        build(childNode, depth + 1);
         node.children.push(childNode);
       }
     }
+  };
 
-    // hide node IFF nothing in its subtree matches
-    const anyMatch = matchesSelf || matchesChild;
-    node.hidden = !anyMatch;
-
-    return anyMatch;
+  // Hide a node IFF nothing in its subtree matches; with no query, nothing is hidden
+  let matched: Set<TreeNode<SidebarModel>> | null = null;
+  const markHidden = (node: TreeNode<SidebarModel>): boolean => {
+    let visible = matched == null || matched.has(node);
+    for (const child of node.children ?? []) {
+      if (markHidden(child)) visible = true;
+    }
+    node.hidden = !visible;
+    return visible;
   };
 
   const root: TreeNode<SidebarModel> = {
@@ -974,15 +997,17 @@ const sidebarTreeAtom = atom<
     depth: 0,
   };
 
-  // Build tree and mark visibility in one pass
   build(root, 1);
+  if (!matcher.isEmpty) {
+    const rows = matcher.matchingRows(sidebarFilterColumns, cells);
+    matched = new Set(Array.from(rows, (i) => leaves[i]!));
+  }
+  matcher.free();
+  markHidden(root);
 
-  const fields: FieldDef[] = [];
+  const fieldValues: Record<string, string[]> = {};
   for (const [name, values] of Object.entries(allFields)) {
-    fields.push({
-      name,
-      values: Array.from(values).filter((v) => v.length < 20),
-    });
+    fieldValues[name] = Array.from(values).filter((v) => v.length < 20);
   }
   const suggestions = Array.from(suggestionFields)
     .sort((a, b) => {
@@ -993,9 +1018,9 @@ const sidebarTreeAtom = atom<
     })
     .map((field) => ({
       field,
-      filterText: formatFieldFilter(field, suggestionValue ?? ""),
+      filterText: formatField(field, suggestionValue ?? ""),
     }));
-  return [root, fields, suggestions] as const;
+  return [root, fieldValues, suggestions] as const;
 });
 
 const sidebarGitStatusByModelIdAtom = atom<Record<string, GitStatus>>((get) => {
diff --git a/apps/yaak-client/components/core/Editor/filter/extension.ts b/apps/yaak-client/components/core/Editor/filter/extension.ts
new file mode 100644
index 000000000..003729fe5
--- /dev/null
+++ b/apps/yaak-client/components/core/Editor/filter/extension.ts
@@ -0,0 +1,88 @@
+import type { Completion, CompletionContext, CompletionResult } from "@codemirror/autocomplete";
+import { autocompletion, startCompletion } from "@codemirror/autocomplete";
+import { highlightingFor } from "@codemirror/language";
+import type { EditorState, Extension } from "@codemirror/state";
+import { RangeSetBuilder } from "@codemirror/state";
+import type { DecorationSet, ViewUpdate } from "@codemirror/view";
+import { Decoration, EditorView, ViewPlugin } from "@codemirror/view";
+import type { Tag } from "@lezer/highlight";
+import { tags as t } from "@lezer/highlight";
+import type { Schema, TokenKind } from "@yaakapp-internal/query";
+import { queryCompletions, queryTokens } from "@yaakapp-internal/query";
+
+export interface QueryFilterOptions {
+  schema: Schema;
+  /** The values seen for each field path, offered after `@field:` */
+  values?: Record<string, string[]>;
+}
+
+const tokenTags: Record<TokenKind, Tag | null> = {
+  field: t.attributeName,
+  value: t.attributeValue,
+  operator: t.operatorKeyword,
+  quote: t.string,
+  paren: t.paren,
+  negation: t.operator,
+  text: null,
+};
+
+function buildDecorations(state: EditorState): DecorationSet {
+  const builder = new RangeSetBuilder<Decoration>();
+  for (const token of queryTokens(state.doc.toString())) {
+    const tag = tokenTags[token.kind];
+    const className = tag == null ? null : highlightingFor(state, [tag]);
+    if (className == null || token.span.start === token.span.end) continue;
+    builder.add(token.span.start, token.span.end, Decoration.mark({ class: className }));
+  }
+  return builder.finish();
+}
+
+const highlight = ViewPlugin.fromClass(
+  class {
+    decorations: DecorationSet;
+
+    constructor(view: EditorView) {
+      this.decorations = buildDecorations(view.state);
+    }
+
+    update(update: ViewUpdate) {
+      if (update.docChanged || update.viewportChanged) {
+        this.decorations = buildDecorations(update.state);
+      }
+    }
+  },
+  { decorations: (v) => v.decorations },
+);
+
+function completionSource({ schema, values }: QueryFilterOptions) {
+  return (ctx: CompletionContext): CompletionResult | null => {
+    const result = queryCompletions(ctx.state.doc.toString(), ctx.pos, schema, values);
+    if (result == null) return null;
+
+    const { start, end } = result.span;
+    // Field names pop up for something typed; on an empty spot, only when asked for
+    const onlyFields = result.items.every((item) => item.kind === "field");
+    if (onlyFields && start === end && !ctx.explicit) return null;
+
+    const options: Completion[] = result.items.map((item) => ({
+      label: item.label,
+      type: item.kind === "field" ? "property" : "constant",
+      info: item.info ?? undefined,
+      apply: (view, _completion, from, to) => {
+        view.dispatch({
+          changes: { from, to, insert: item.insert },
+          selection: { anchor: from + item.insert.length },
+        });
+        if (item.reopen) startCompletion(view);
+      },
+    }));
+
+    // The engine has already filtered and ranked the options
+    return { from: start, to: end, options, filter: false };
+  };
+}
+
+/** Highlighting and completion for a search box written in Yaak's query language */
+export function queryFilter(options: QueryFilterOptions): Extension {
+  return [highlight, autocompletion({ override: [completionSource(options)] })];
+}
diff --git a/apps/yaak-client/components/core/Editor/filter/format.test.ts b/apps/yaak-client/components/core/Editor/filter/format.test.ts
deleted file mode 100644
index 07ba7afe6..000000000
--- a/apps/yaak-client/components/core/Editor/filter/format.test.ts
+++ /dev/null
@@ -1,43 +0,0 @@
-import { describe, expect, test } from "vite-plus/test";
-import { formatFieldFilter } from "./format";
-import { evaluate, parseQuery } from "./query";
-
-function matchesFormattedUrl(value: string) {
-  return evaluate(parseQuery(formatFieldFilter("url", value)), {
-    fields: { url: value },
-  });
-}
-
-describe("formatFieldFilter", () => {
-  test("keeps URL-like values bare", () => {
-    expect(formatFieldFilter("url", "yaak.app/foo-bar")).toBe("@url:yaak.app/foo-bar");
-    expect(matchesFormattedUrl("yaak.app/foo-bar")).toBe(true);
-  });
-
-  test("keeps non-syntax punctuation bare", () => {
-    expect(formatFieldFilter("url", "yaa$&#*@tsrna(*)")).toBe("@url:yaa$&#*@tsrna(*)");
-    expect(matchesFormattedUrl("yaa$&#*@tsrna(*)")).toBe(true);
-  });
-
-  test("keeps values that start with an operator token bare", () => {
-    expect(formatFieldFilter("url", "-foo")).toBe("@url:-foo");
-    expect(matchesFormattedUrl("-foo")).toBe(true);
-  });
-
-  test("keeps boolean operator words bare", () => {
-    expect(formatFieldFilter("url", "AND")).toBe("@url:AND");
-    expect(formatFieldFilter("url", "or")).toBe("@url:or");
-    expect(formatFieldFilter("url", "Not")).toBe("@url:Not");
-    expect(matchesFormattedUrl("AND")).toBe(true);
-  });
-
-  test("escapes quoted values", () => {
-    expect(formatFieldFilter("url", 'say "hi"')).toBe('@url:"say \\"hi\\""');
-    expect(matchesFormattedUrl('say "hi"')).toBe(true);
-  });
-
-  test("quotes values that start with a quote", () => {
-    expect(formatFieldFilter("url", '"hi"')).toBe('@url:"\\"hi\\""');
-    expect(matchesFormattedUrl('"hi"')).toBe(true);
-  });
-});
diff --git a/apps/yaak-client/components/core/Editor/filter/format.ts b/apps/yaak-client/components/core/Editor/filter/format.ts
deleted file mode 100644
index f1fc7215e..000000000
--- a/apps/yaak-client/components/core/Editor/filter/format.ts
+++ /dev/null
@@ -1,7 +0,0 @@
-const bareFieldValue = /^[^\s"]\S*$/;
-
-export function formatFieldFilter(field: string, value: string) {
-  const escapedValue = value.replace(/\\/g, "\\\\").replace(/"/g, '\\"');
-  const filterValue = bareFieldValue.test(value) ? value : `"${escapedValue}"`;
-  return `@${field}:${filterValue}`;
-}
diff --git a/apps/yaak-client/components/core/Editor/filter/query.ts b/apps/yaak-client/components/core/Editor/filter/query.ts
deleted file mode 100644
index 12bf4a4bb..000000000
--- a/apps/yaak-client/components/core/Editor/filter/query.ts
+++ /dev/null
@@ -1,308 +0,0 @@
-// query.ts
-// A tiny query language parser with NOT/AND/OR, parentheses, phrases, negation, and field:value.
-
-import { fuzzyMatch } from "fuzzbunny";
-/////////////////////////
-// AST
-/////////////////////////
-
-export type Ast =
-  | { type: "Term"; value: string } // foo
-  | { type: "Phrase"; value: string } // "hi there"
-  | { type: "Field"; field: string; value: string } // method:POST or title:"exact phrase"
-  | { type: "Not"; node: Ast } // -foo or NOT foo
-  | { type: "And"; left: Ast; right: Ast } // a AND b
-  | { type: "Or"; left: Ast; right: Ast }; // a OR b
-
-/////////////////////////
-// Tokenizer
-/////////////////////////
-type Tok =
-  | { kind: "LPAREN" }
-  | { kind: "RPAREN" }
-  | { kind: "AND" }
-  | { kind: "OR" }
-  | { kind: "NOT" } // explicit NOT
-  | { kind: "MINUS" } // unary minus before term/phrase/paren group
-  | { kind: "COLON" }
-  | { kind: "WORD"; text: string } // bareword (unquoted)
-  | { kind: "PHRASE"; text: string } // "quoted phrase"
-  | { kind: "EOF" };
-
-const isSpace = (c: string) => /\s/.test(c);
-const isWordStart = (c: string) => c !== "" && !isSpace(c) && !/[():"@-]/.test(c);
-const isWordChar = (c: string) => c !== "" && !isSpace(c) && !/[():"@]/.test(c);
-
-export function tokenize(input: string): Tok[] {
-  const toks: Tok[] = [];
-  let i = 0;
-  const n = input.length;
-
-  const peek = () => input[i] ?? "";
-  const advance = () => input[i++];
-
-  const readWord = () => {
-    let s = "";
-    while (i < n && isWordChar(peek())) s += advance();
-    return s;
-  };
-
-  const readFieldValue = () => {
-    let s = "";
-    while (i < n && !isSpace(peek())) s += advance();
-    return s;
-  };
-
-  const readPhrase = () => {
-    // assumes current char is opening quote
-    advance(); // consume opening "
-    let s = "";
-    while (i < n) {
-      const c = advance();
-      if (c === `"`) break;
-      if (c === "\\" && i < n) {
-        // escape \" and \\ (simple)
-        const next = advance();
-        s += next;
-      } else {
-        s += c;
-      }
-    }
-    return s;
-  };
-
-  while (i < n) {
-    const c = peek();
-
-    if (isSpace(c)) {
-      i++;
-      continue;
-    }
-
-    if (c === "(") {
-      toks.push({ kind: "LPAREN" });
-      i++;
-      continue;
-    }
-    if (c === ")") {
-      toks.push({ kind: "RPAREN" });
-      i++;
-      continue;
-    }
-    if (c === ":") {
-      toks.push({ kind: "COLON" });
-      i++;
-      if (peek() && !isSpace(peek()) && peek() !== `"`) {
-        toks.push({ kind: "WORD", text: readFieldValue() });
-      }
-      continue;
-    }
-    if (c === `"`) {
-      const text = readPhrase();
-      toks.push({ kind: "PHRASE", text });
-      continue;
-    }
-    if (c === "-") {
-      toks.push({ kind: "MINUS" });
-      i++;
-      continue;
-    }
-
-    // WORD / AND / OR / NOT
-    if (isWordStart(c)) {
-      const w = readWord();
-      const upper = w.toUpperCase();
-      if (upper === "AND") toks.push({ kind: "AND" });
-      else if (upper === "OR") toks.push({ kind: "OR" });
-      else if (upper === "NOT") toks.push({ kind: "NOT" });
-      else toks.push({ kind: "WORD", text: w });
-      continue;
-    }
-
-    // Unknown char—skip to be forgiving
-    i++;
-  }
-
-  toks.push({ kind: "EOF" });
-  return toks;
-}
-
-class Parser {
-  private i = 0;
-  constructor(private toks: Tok[]) {}
-
-  private peek(): Tok {
-    return this.toks[this.i] ?? { kind: "EOF" };
-  }
-  private advance(): Tok {
-    return this.toks[this.i++] ?? { kind: "EOF" };
-  }
-  private at(kind: Tok["kind"]) {
-    return this.peek().kind === kind;
-  }
-
-  // Top-level: parse OR-precedence chain, allowing implicit AND.
-  parse(): Ast | null {
-    if (this.at("EOF")) return null;
-    const expr = this.parseOr();
-    if (!this.at("EOF")) {
-      // Optionally, consume remaining tokens or throw
-    }
-    return expr;
-  }
-
-  // Precedence: NOT (highest), AND, OR (lowest)
-  private parseOr(): Ast {
-    let node = this.parseAnd();
-    while (this.at("OR")) {
-      this.advance();
-      const rhs = this.parseAnd();
-      node = { type: "Or", left: node, right: rhs };
-    }
-    return node;
-  }
-
-  private parseAnd(): Ast {
-    let node = this.parseUnary();
-    // Implicit AND: if next token starts a primary, treat as AND.
-    while (this.at("AND") || this.startsPrimary()) {
-      if (this.at("AND")) this.advance();
-      const rhs = this.parseUnary();
-      node = { type: "And", left: node, right: rhs };
-    }
-    return node;
-  }
-
-  private parseUnary(): Ast {
-    if (this.at("NOT") || this.at("MINUS")) {
-      this.advance();
-      const node = this.parseUnary();
-      return { type: "Not", node };
-    }
-    return this.parsePrimaryOrField();
-  }
-
-  private startsPrimary(): boolean {
-    const k = this.peek().kind;
-    return k === "WORD" || k === "PHRASE" || k === "LPAREN" || k === "MINUS" || k === "NOT";
-  }
-
-  private parsePrimaryOrField(): Ast {
-    // Parenthesized group
-    if (this.at("LPAREN")) {
-      this.advance();
-      const inside = this.parseOr();
-      // if (!this.at('RPAREN')) throw new Error("Missing closing ')'");
-      this.advance();
-      return inside;
-    }
-
-    // Phrase
-    if (this.at("PHRASE")) {
-      const t = this.advance() as Extract<Tok, { kind: "PHRASE" }>;
-      return { type: "Phrase", value: t.text };
-    }
-
-    // Field or bare word
-    if (this.at("WORD")) {
-      const wordTok = this.advance() as Extract<Tok, { kind: "WORD" }>;
-
-      if (this.at("COLON")) {
-        // field:value or field:"phrase"
-        this.advance(); // :
-        let value: string;
-        if (this.at("PHRASE")) {
-          const p = this.advance() as Extract<Tok, { kind: "PHRASE" }>;
-          value = p.text;
-        } else if (this.at("WORD")) {
-          const w = this.advance() as Extract<Tok, { kind: "WORD" }>;
-          value = w.text;
-        } else {
-          // Anything else after colon is treated literally as a single Term token.
-          const t = this.advance();
-          value = tokText(t);
-        }
-        return { type: "Field", field: wordTok.text, value };
-      }
-
-      // plain term
-      return { type: "Term", value: wordTok.text };
-    }
-
-    const w = this.advance() as Extract<Tok, { kind: "WORD" }>;
-    return { type: "Phrase", value: "text" in w ? w.text : "" };
-  }
-}
-
-function tokText(t: Tok): string {
-  if ("text" in t) return t.text;
-
-  switch (t.kind) {
-    case "COLON":
-      return ":";
-    case "LPAREN":
-      return "(";
-    case "RPAREN":
-      return ")";
-    default:
-      return "";
-  }
-}
-
-export function parseQuery(q: string): Ast | null {
-  if (q.trim() === "") return null;
-  const toks = tokenize(q);
-  const parser = new Parser(toks);
-  return parser.parse();
-}
-
-export type Doc = {
-  text?: string;
-  fields?: Record<string, unknown>;
-};
-
-type Technique = "substring" | "fuzzy" | "strict";
-
-function includes(hay: string | undefined, needle: string, technique: Technique): boolean {
-  if (!hay || !needle) return false;
-  if (technique === "strict") return hay === needle;
-  if (technique === "fuzzy") return !!fuzzyMatch(hay, needle);
-  return hay.indexOf(needle) !== -1;
-}
-
-export function evaluate(ast: Ast | null, doc: Doc): boolean {
-  if (!ast) return true; // Match everything if no query is provided
-
-  const text = (doc.text ?? "").toLowerCase();
-  const fieldsNorm: Record<string, string[]> = {};
-
-  for (const [k, v] of Object.entries(doc.fields ?? {})) {
-    if (!(typeof v === "string" || Array.isArray(v))) continue;
-    fieldsNorm[k.toLowerCase()] = Array.isArray(v)
-      ? v.filter((v) => typeof v === "string").map((s) => s.toLowerCase())
-      : [String(v ?? "").toLowerCase()];
-  }
-
-  const evalNode = (node: Ast): boolean => {
-    switch (node.type) {
-      case "Term":
-        return includes(text, node.value.toLowerCase(), "fuzzy");
-      case "Phrase":
-        // Quoted phrases match exactly
-        return includes(text, node.value.toLowerCase(), "substring");
-      case "Field": {
-        const vals = fieldsNorm[node.field.toLowerCase()] ?? [];
-        if (vals.length === 0) return false;
-        return vals.some((v) => includes(v, node.value.toLowerCase(), "substring"));
-      }
-      case "Not":
-        return !evalNode(node.node);
-      case "And":
-        return evalNode(node.left) && evalNode(node.right);
-      case "Or":
-        return evalNode(node.left) || evalNode(node.right);
-    }
-  };
-
-  return evalNode(ast);
-}
diff --git a/crates/yaak-query/Cargo.toml b/crates/yaak-query/Cargo.toml
new file mode 100644
index 000000000..cd1813563
--- /dev/null
+++ b/crates/yaak-query/Cargo.toml
@@ -0,0 +1,37 @@
+[package]
+name = "yaak-query"
+version = "0.1.0"
+edition = "2024"
+publish = false
+
+# The search query language shared by Yaak's sidebar and Yaak Proxy: grammar,
+# parser, schema, evaluator, completion and highlighting. Pure Rust, with a
+# WebAssembly build (`@yaakapp-internal/query`) for the frontends.
+
+[package.metadata.wasm-pack.profile.release]
+wasm-opt = false # Matches yaak-templates; wasm-opt has caused errors in CI
+
+[lib]
+crate-type = ["cdylib", "rlib"]
+
+[features]
+default = ["wasm"]
+# The `#[wasm_bindgen]` exports that make up the @yaakapp-internal/query
+# package. They only exist when compiling for wasm32, so native builds and
+# `cargo test` never see them. Off for crates that link this one into their own
+# wasm module and do not want these re-exported from theirs.
+wasm = []
+
+[dependencies]
+serde = { workspace = true, features = ["derive"] }
+
+# Everything that only exists for the WebAssembly package is target-scoped, so
+# a native consumer (the proxy, the CLI) links none of it.
+[target.'cfg(target_arch = "wasm32")'.dependencies]
+serde-wasm-bindgen = "0.6.5"
+serde_json = { workspace = true }
+wasm-bindgen = { version = "0.2.100", features = ["serde-serialize"] }
+
+[dev-dependencies]
+serde_json = { workspace = true }
+ts-rs = { workspace = true }
diff --git a/crates/yaak-query/bindings/query.ts b/crates/yaak-query/bindings/query.ts
new file mode 100644
index 000000000..a9efad919
--- /dev/null
+++ b/crates/yaak-query/bindings/query.ts
@@ -0,0 +1,128 @@
+// This file was generated by [ts-rs](https://github.com/Aleph-Alpha/ts-rs). Do not edit this file manually.
+
+export type CompareOp = ">" | ">=" | "<" | "<=";
+
+export type Completion = { 
+/**
+ * What to show: a field's path, or a value as it is.
+ */
+label: string, 
+/**
+ * What replaces the span: `@method:`, `@header.`, or a value quoted when it needs to be.
+ */
+insert: string, kind: CompletionKind, info: string | null, 
+/**
+ * The term is not finished after inserting (a field still needs its value, a map field its
+ * key), so completion should open again.
+ */
+reopen: boolean, };
+
+export type CompletionKind = "field" | "value";
+
+/**
+ * Completions for the term at the cursor, all replacing `span`.
+ */
+export type Completions = { span: Span, items: Array<Completion>, };
+
+/**
+ * Something in a query the schema does not accept. The query still parses and runs; the
+ * offending term matches nothing.
+ */
+export type Diagnostic = { "type": "unknownField", path: string, span: Span, } | { "type": "invalidValue", path: string, span: Span, };
+
+export type DocValue = boolean | number | string;
+
+/**
+ * One value or several.
+ */
+export type DocValues = DocValue | Array<DocValue>;
+
+/**
+ * What a query is matched against: free text for bare words and phrases, and field values by
+ * lowercase dotted path. A map field's values sit under `name.key` (`header.x-cache`).
+ */
+export type Document = { text?: string, fields?: { [key in string]?: DocValues }, };
+
+/**
+ * One field: `@name:value`.
+ */
+export type FieldDef = { 
+/**
+ * The field's path, dotted when nested: `method`, `req.header`.
+ */
+name: string, kind: FieldKind, 
+/**
+ * A map of named values, such as headers: `@header.<key>:value` is valid for any key, and
+ * the key is matched against the document's keys (case-insensitively).
+ */
+map?: boolean, 
+/**
+ * A short note shown with the field in completions.
+ */
+info?: string, };
+
+export type FieldKind = { "type": "text" } | { "type": "number", units: Array<Unit>, } | { "type": "flag" };
+
+/**
+ * What a field term's value asks for. The shape is decided by the parser alone; the schema
+ * decides what it means for a given field.
+ */
+export type FieldValue = { "type": "text", text: string, quoted: boolean, } | { "type": "wildcard", pattern: string, } | { "type": "exists" } | { "type": "compare", op: CompareOp, number: NumberLiteral, } | { "type": "range", low: NumberLiteral, high: NumberLiteral, };
+
+/**
+ * How a bare word matches the document's free text. A quoted phrase is always a substring.
+ */
+export type FreeText = "fuzzy" | "substring";
+
+/**
+ * A node of the parsed query.
+ */
+export type Node = { "type": "And", children: Array<Node>, span: Span, } | { "type": "Or", children: Array<Node>, span: Span, } | { "type": "Not", node: Node, span: Span, } | { "type": "Term", text: string, span: Span, } | { "type": "Phrase", text: string, span: Span, } | { "type": "Field", 
+/**
+ * The dotted path, one entry per identifier: `@req.header.accept` is
+ * `["req", "header", "accept"]`.
+ */
+path: Array<string>, value: FieldValue, span: Span, 
+/**
+ * `@path:`, from the `@` through the colon.
+ */
+pathSpan: Span, 
+/**
+ * The value as written, quotes included. Empty (and at the end of the field) when
+ * nothing follows the colon.
+ */
+valueSpan: Span, };
+
+/**
+ * A number as written in a comparison or range, with its unit when it has one (`500ms`,
+ * `2.5mb`). The schema says which units a field accepts and what they are worth.
+ */
+export type NumberLiteral = { value: number, unit: string | null, 
+/**
+ * The number as it appears in the source, unit included.
+ */
+raw: string, };
+
+/**
+ * A consumer's searchable fields.
+ */
+export type Schema = { fields: Array<FieldDef>, 
+/**
+ * Substring when left out.
+ */
+freeText?: FreeText, };
+
+/**
+ * A half-open byte range `start..end` in the query source.
+ */
+export type Span = { start: number, end: number, };
+
+export type Token = { kind: TokenKind, span: Span, };
+
+export type TokenKind = "field" | "value" | "operator" | "quote" | "paren" | "negation" | "text";
+
+/**
+ * A unit a number may be written in, and how many of the field's base unit it is worth:
+ * `{ name: "s", factor: 1000 }` for a field counted in milliseconds.
+ */
+export type Unit = { name: string, factor: number, };
diff --git a/crates/yaak-query/build-wasm.cjs b/crates/yaak-query/build-wasm.cjs
new file mode 100644
index 000000000..22a6b65b3
--- /dev/null
+++ b/crates/yaak-query/build-wasm.cjs
@@ -0,0 +1,40 @@
+const { execSync } = require("node:child_process");
+const fs = require("node:fs");
+const os = require("node:os");
+const path = require("node:path");
+
+if (process.env.SKIP_WASM_BUILD === "1") {
+  console.log("Skipping wasm-pack build (SKIP_WASM_BUILD=1)");
+  return;
+}
+
+// Remap machine-specific paths that rustc embeds into the binary (panic
+// location strings), so builds are reproducible across machines
+const sysroot = execSync("rustc --print sysroot").toString().trim();
+const cargoHome = process.env.CARGO_HOME ?? path.join(os.homedir(), ".cargo");
+
+execSync("wasm-pack build --target bundler", {
+  stdio: "inherit",
+  env: {
+    ...process.env,
+    CARGO_BUILD_TARGET: "wasm32-unknown-unknown",
+    RUSTFLAGS: `--remap-path-prefix=${cargoHome}=/cargo --remap-path-prefix=${sysroot}=/rustc`,
+  },
+});
+
+// Rewrite the generated entry to use Vite's ?init import style instead of
+// the ES Module Integration style that wasm-pack generates, which Vite/rolldown
+// does not support in production builds.
+const entry = path.join(__dirname, "pkg", "yaak_query.js");
+fs.writeFileSync(
+  entry,
+  [
+    'import init from "./yaak_query_bg.wasm?init";',
+    'export * from "./yaak_query_bg.js";',
+    'import * as bg from "./yaak_query_bg.js";',
+    'const instance = await init({ "./yaak_query_bg.js": bg });',
+    "bg.__wbg_set_wasm(instance.exports);",
+    "instance.exports.__wbindgen_start();",
+    "",
+  ].join("\n"),
+);
diff --git a/crates/yaak-query/index.ts b/crates/yaak-query/index.ts
new file mode 100644
index 000000000..14ca5657a
--- /dev/null
+++ b/crates/yaak-query/index.ts
@@ -0,0 +1,85 @@
+export * from "./bindings/query";
+import type { Completions, Diagnostic, Document, Node, Schema, Token } from "./bindings/query";
+import {
+  QueryMatcher as WasmQueryMatcher,
+  parse_query,
+  query_completions,
+  query_diagnostics,
+  query_format_field,
+  query_tokens,
+} from "./pkg";
+
+// Every span and offset here counts UTF-16 code units, like JavaScript strings do.
+
+/** The parsed tree, or `null` when the query has no terms */
+export function parseQuery(input: string): Node | null {
+  return parse_query(input) as Node | null;
+}
+
+/** Highlighting spans, in order */
+export function queryTokens(input: string): Token[] {
+  return query_tokens(input) as Token[];
+}
+
+/**
+ * What can complete the term at `cursor`: field names, or after `@field:` the values given for
+ * that field in `values`
+ */
+export function queryCompletions(
+  input: string,
+  cursor: number,
+  schema: Schema,
+  values?: Record<string, string[]>,
+): Completions | null {
+  return query_completions(input, cursor, schema, values ?? {}) as Completions | null;
+}
+
+/** The terms the schema cannot run: unknown fields and values that cannot apply */
+export function queryDiagnostics(input: string, schema: Schema): Diagnostic[] {
+  return query_diagnostics(input, schema) as Diagnostic[];
+}
+
+/** `@field:value` (or `-@field:value`), with the value quoted when it needs to be */
+export function formatField(field: string, value: string, negated = false): string {
+  return query_format_field(field, value, negated);
+}
+
+/** The length that marks a cell with no value (`Table::NO_VALUE` in the crate) */
+const NO_VALUE = 0xffffffff;
+
+/** A query parsed once, to match against many documents */
+export class QueryMatcher {
+  readonly #inner: WasmQueryMatcher;
+
+  constructor(input: string, schema: Schema) {
+    this.#inner = new WasmQueryMatcher(input, schema);
+  }
+
+  /** The query has no terms, so it matches everything */
+  get isEmpty(): boolean {
+    return this.#inner.isEmpty;
+  }
+
+  matches(doc: Document): boolean {
+    return this.#inner.matches(JSON.stringify(doc));
+  }
+
+  /**
+   * The indexes of the matching rows, for many documents in one call. `cells` is flat: for
+   * each row, its free text, then a value (or `null` when it has none) for each of `columns`
+   */
+  matchingRows(columns: string[], cells: (string | null)[]): Uint32Array {
+    // Packed as one string plus each cell's length, which crosses into wasm far faster than
+    // JSON: nothing to escape or parse, and the cells are read in place
+    const lengths = new Uint32Array(cells.length);
+    for (let i = 0; i < cells.length; i++) {
+      lengths[i] = cells[i]?.length ?? NO_VALUE;
+    }
+    return this.#inner.matchingRows(columns, cells.join(""), lengths);
+  }
+
+  /** Frees the parsed query now rather than when it is garbage collected */
+  free() {
+    this.#inner.free();
+  }
+}
diff --git a/crates/yaak-query/package.json b/crates/yaak-query/package.json
new file mode 100644
index 000000000..da9e0b86a
--- /dev/null
+++ b/crates/yaak-query/package.json
@@ -0,0 +1,15 @@
+{
+  "name": "@yaakapp-internal/query",
+  "version": "1.0.0",
+  "private": true,
+  "main": "index.ts",
+  "scripts": {
+    "bootstrap": "npm run build",
+    "build": "run-s build:*",
+    "build:pack": "node build-wasm.cjs",
+    "build:clean": "rimraf ./pkg/.gitignore"
+  },
+  "devDependencies": {
+    "rimraf": "^6.1.2"
+  }
+}
diff --git a/crates/yaak-query/pkg/package.json b/crates/yaak-query/pkg/package.json
new file mode 100644
index 000000000..a51f42d53
--- /dev/null
+++ b/crates/yaak-query/pkg/package.json
@@ -0,0 +1,17 @@
+{
+  "name": "yaak-query",
+  "type": "module",
+  "version": "0.1.0",
+  "files": [
+    "yaak_query_bg.wasm",
+    "yaak_query.js",
+    "yaak_query_bg.js",
+    "yaak_query.d.ts"
+  ],
+  "main": "yaak_query.js",
+  "types": "yaak_query.d.ts",
+  "sideEffects": [
+    "./yaak_query.js",
+    "./snippets/*"
+  ]
+}
\ No newline at end of file
diff --git a/crates/yaak-query/pkg/yaak_query.d.ts b/crates/yaak-query/pkg/yaak_query.d.ts
new file mode 100644
index 000000000..d326c7931
--- /dev/null
+++ b/crates/yaak-query/pkg/yaak_query.d.ts
@@ -0,0 +1,49 @@
+/* tslint:disable */
+/* eslint-disable */
+
+/**
+ * A query parsed and resolved once, then matched against many documents.
+ */
+export class QueryMatcher {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Whether a `Document` (`{ text, fields }`), as JSON, matches. JSON crosses into wasm far
+     * faster than walking a JS object property by property.
+     */
+    matches(doc_json: string): boolean;
+    /**
+     * The indexes of the matching rows, for many documents in one call. The cells (per row,
+     * its free text, then one per column) are packed as in [`Table::from_packed`]: joined into
+     * one string, with each one's UTF-16 length, or `Table::NO_VALUE` for no value.
+     */
+    matchingRows(columns: string[], joined: string, lengths: Uint32Array): Uint32Array;
+    constructor(input: string, schema: any);
+    /**
+     * Whether the query has no terms, so it matches everything.
+     */
+    readonly isEmpty: boolean;
+}
+
+/**
+ * The parsed tree (`Node | null`), with UTF-16 spans.
+ */
+export function parse_query(input: string): any;
+
+/**
+ * `Completions | null` at UTF-16 offset `cursor`. `values` maps field paths to the values seen
+ * for them (`Record<string, string[]>`).
+ */
+export function query_completions(input: string, cursor: number, schema: any, values: any): any;
+
+/**
+ * The terms the schema cannot run (`Diagnostic[]`), with UTF-16 spans.
+ */
+export function query_diagnostics(input: string, schema: any): any;
+
+export function query_format_field(field: string, value: string, negated: boolean): string;
+
+/**
+ * Highlighting spans (`Token[]`), in UTF-16 offsets.
+ */
+export function query_tokens(input: string): any;
diff --git a/crates/yaak-query/pkg/yaak_query.js b/crates/yaak-query/pkg/yaak_query.js
new file mode 100644
index 000000000..d3d8a23f4
--- /dev/null
+++ b/crates/yaak-query/pkg/yaak_query.js
@@ -0,0 +1,6 @@
+import init from "./yaak_query_bg.wasm?init";
+export * from "./yaak_query_bg.js";
+import * as bg from "./yaak_query_bg.js";
+const instance = await init({ "./yaak_query_bg.js": bg });
+bg.__wbg_set_wasm(instance.exports);
+instance.exports.__wbindgen_start();
diff --git a/crates/yaak-query/pkg/yaak_query_bg.js b/crates/yaak-query/pkg/yaak_query_bg.js
new file mode 100644
index 000000000..8b94229f4
--- /dev/null
+++ b/crates/yaak-query/pkg/yaak_query_bg.js
@@ -0,0 +1,609 @@
+/**
+ * A query parsed and resolved once, then matched against many documents.
+ */
+export class QueryMatcher {
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        QueryMatcherFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_querymatcher_free(ptr, 0);
+    }
+    /**
+     * Whether the query has no terms, so it matches everything.
+     * @returns {boolean}
+     */
+    get isEmpty() {
+        const ret = wasm.querymatcher_isEmpty(this.__wbg_ptr);
+        return ret !== 0;
+    }
+    /**
+     * Whether a `Document` (`{ text, fields }`), as JSON, matches. JSON crosses into wasm far
+     * faster than walking a JS object property by property.
+     * @param {string} doc_json
+     * @returns {boolean}
+     */
+    matches(doc_json) {
+        const ptr0 = passStringToWasm0(doc_json, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.querymatcher_matches(this.__wbg_ptr, ptr0, len0);
+        if (ret[2]) {
+            throw takeFromExternrefTable0(ret[1]);
+        }
+        return ret[0] !== 0;
+    }
+    /**
+     * The indexes of the matching rows, for many documents in one call. The cells (per row,
+     * its free text, then one per column) are packed as in [`Table::from_packed`]: joined into
+     * one string, with each one's UTF-16 length, or `Table::NO_VALUE` for no value.
+     * @param {string[]} columns
+     * @param {string} joined
+     * @param {Uint32Array} lengths
+     * @returns {Uint32Array}
+     */
+    matchingRows(columns, joined, lengths) {
+        const ptr0 = passArrayJsValueToWasm0(columns, wasm.__wbindgen_malloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ptr1 = passStringToWasm0(joined, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len1 = WASM_VECTOR_LEN;
+        const ptr2 = passArray32ToWasm0(lengths, wasm.__wbindgen_malloc);
+        const len2 = WASM_VECTOR_LEN;
+        const ret = wasm.querymatcher_matchingRows(this.__wbg_ptr, ptr0, len0, ptr1, len1, ptr2, len2);
+        if (ret[3]) {
+            throw takeFromExternrefTable0(ret[2]);
+        }
+        var v4 = getArrayU32FromWasm0(ret[0], ret[1]).slice();
+        wasm.__wbindgen_free(ret[0], ret[1] * 4, 4);
+        return v4;
+    }
+    /**
+     * @param {string} input
+     * @param {any} schema
+     */
+    constructor(input, schema) {
+        const ptr0 = passStringToWasm0(input, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ret = wasm.querymatcher_new(ptr0, len0, schema);
+        if (ret[2]) {
+            throw takeFromExternrefTable0(ret[1]);
+        }
+        this.__wbg_ptr = ret[0];
+        QueryMatcherFinalization.register(this, this.__wbg_ptr, this);
+        return this;
+    }
+}
+if (Symbol.dispose) QueryMatcher.prototype[Symbol.dispose] = QueryMatcher.prototype.free;
+
+/**
+ * The parsed tree (`Node | null`), with UTF-16 spans.
+ * @param {string} input
+ * @returns {any}
+ */
+export function parse_query(input) {
+    const ptr0 = passStringToWasm0(input, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len0 = WASM_VECTOR_LEN;
+    const ret = wasm.parse_query(ptr0, len0);
+    if (ret[2]) {
+        throw takeFromExternrefTable0(ret[1]);
+    }
+    return takeFromExternrefTable0(ret[0]);
+}
+
+/**
+ * `Completions | null` at UTF-16 offset `cursor`. `values` maps field paths to the values seen
+ * for them (`Record<string, string[]>`).
+ * @param {string} input
+ * @param {number} cursor
+ * @param {any} schema
+ * @param {any} values
+ * @returns {any}
+ */
+export function query_completions(input, cursor, schema, values) {
+    const ptr0 = passStringToWasm0(input, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len0 = WASM_VECTOR_LEN;
+    const ret = wasm.query_completions(ptr0, len0, cursor, schema, values);
+    if (ret[2]) {
+        throw takeFromExternrefTable0(ret[1]);
+    }
+    return takeFromExternrefTable0(ret[0]);
+}
+
+/**
+ * The terms the schema cannot run (`Diagnostic[]`), with UTF-16 spans.
+ * @param {string} input
+ * @param {any} schema
+ * @returns {any}
+ */
+export function query_diagnostics(input, schema) {
+    const ptr0 = passStringToWasm0(input, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len0 = WASM_VECTOR_LEN;
+    const ret = wasm.query_diagnostics(ptr0, len0, schema);
+    if (ret[2]) {
+        throw takeFromExternrefTable0(ret[1]);
+    }
+    return takeFromExternrefTable0(ret[0]);
+}
+
+/**
+ * @param {string} field
+ * @param {string} value
+ * @param {boolean} negated
+ * @returns {string}
+ */
+export function query_format_field(field, value, negated) {
+    let deferred3_0;
+    let deferred3_1;
+    try {
+        const ptr0 = passStringToWasm0(field, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len0 = WASM_VECTOR_LEN;
+        const ptr1 = passStringToWasm0(value, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+        const len1 = WASM_VECTOR_LEN;
+        const ret = wasm.query_format_field(ptr0, len0, ptr1, len1, negated);
+        deferred3_0 = ret[0];
+        deferred3_1 = ret[1];
+        return getStringFromWasm0(ret[0], ret[1]);
+    } finally {
+        wasm.__wbindgen_free(deferred3_0, deferred3_1, 1);
+    }
+}
+
+/**
+ * Highlighting spans (`Token[]`), in UTF-16 offsets.
+ * @param {string} input
+ * @returns {any}
+ */
+export function query_tokens(input) {
+    const ptr0 = passStringToWasm0(input, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len0 = WASM_VECTOR_LEN;
+    const ret = wasm.query_tokens(ptr0, len0);
+    if (ret[2]) {
+        throw takeFromExternrefTable0(ret[1]);
+    }
+    return takeFromExternrefTable0(ret[0]);
+}
+export function __wbg_Error_bce6d499ff0a4aff(arg0, arg1) {
+    const ret = Error(getStringFromWasm0(arg0, arg1));
+    return ret;
+}
+export function __wbg_String_8564e559799eccda(arg0, arg1) {
+    const ret = String(arg1);
+    const ptr1 = passStringToWasm0(ret, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len1 = WASM_VECTOR_LEN;
+    getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+    getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+}
+export function __wbg___wbindgen_bigint_get_as_i64_410e28c7b761ad83(arg0, arg1) {
+    const v = arg1;
+    const ret = typeof(v) === 'bigint' ? v : undefined;
+    getDataViewMemory0().setBigInt64(arg0 + 8 * 1, isLikeNone(ret) ? BigInt(0) : ret, true);
+    getDataViewMemory0().setInt32(arg0 + 4 * 0, !isLikeNone(ret), true);
+}
+export function __wbg___wbindgen_boolean_get_2304fb8c853028c8(arg0) {
+    const v = arg0;
+    const ret = typeof(v) === 'boolean' ? v : undefined;
+    return isLikeNone(ret) ? 0xFFFFFF : ret ? 1 : 0;
+}
+export function __wbg___wbindgen_debug_string_edece8177ad01481(arg0, arg1) {
+    const ret = debugString(arg1);
+    const ptr1 = passStringToWasm0(ret, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    const len1 = WASM_VECTOR_LEN;
+    getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+    getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+}
+export function __wbg___wbindgen_in_07056af4f902c445(arg0, arg1) {
+    const ret = arg0 in arg1;
+    return ret;
+}
+export function __wbg___wbindgen_is_bigint_aeae3893f30ed54e(arg0) {
+    const ret = typeof(arg0) === 'bigint';
+    return ret;
+}
+export function __wbg___wbindgen_is_function_5cd60d5cf78b4eef(arg0) {
+    const ret = typeof(arg0) === 'function';
+    return ret;
+}
+export function __wbg___wbindgen_is_null_2042690d351e14f0(arg0) {
+    const ret = arg0 === null;
+    return ret;
+}
+export function __wbg___wbindgen_is_object_b4593df85baada48(arg0) {
+    const val = arg0;
+    const ret = typeof(val) === 'object' && val !== null;
+    return ret;
+}
+export function __wbg___wbindgen_is_string_dde0fd9020db4434(arg0) {
+    const ret = typeof(arg0) === 'string';
+    return ret;
+}
+export function __wbg___wbindgen_is_undefined_35bb9f4c7fd651d5(arg0) {
+    const ret = arg0 === undefined;
+    return ret;
+}
+export function __wbg___wbindgen_jsval_eq_c0ed08b3e0f393b9(arg0, arg1) {
+    const ret = arg0 === arg1;
+    return ret;
+}
+export function __wbg___wbindgen_jsval_loose_eq_0ad77b7717db155c(arg0, arg1) {
+    const ret = arg0 == arg1;
+    return ret;
+}
+export function __wbg___wbindgen_number_get_f73a1244370fcc2c(arg0, arg1) {
+    const obj = arg1;
+    const ret = typeof(obj) === 'number' ? obj : undefined;
+    getDataViewMemory0().setFloat64(arg0 + 8 * 1, isLikeNone(ret) ? 0 : ret, true);
+    getDataViewMemory0().setInt32(arg0 + 4 * 0, !isLikeNone(ret), true);
+}
+export function __wbg___wbindgen_string_get_d109740c0d18f4d7(arg0, arg1) {
+    const obj = arg1;
+    const ret = typeof(obj) === 'string' ? obj : undefined;
+    var ptr1 = isLikeNone(ret) ? 0 : passStringToWasm0(ret, wasm.__wbindgen_malloc, wasm.__wbindgen_realloc);
+    var len1 = WASM_VECTOR_LEN;
+    getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+    getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+}
+export function __wbg___wbindgen_throw_9c31b086c2b26051(arg0, arg1) {
+    throw new Error(getStringFromWasm0(arg0, arg1));
+}
+export function __wbg_call_13665d9f14390edc() { return handleError(function (arg0, arg1) {
+    const ret = arg0.call(arg1);
+    return ret;
+}, arguments); }
+export function __wbg_done_54b8da57023b7ed2(arg0) {
+    const ret = arg0.done;
+    return ret;
+}
+export function __wbg_entries_564a7e8b1e54ede5(arg0) {
+    const ret = Object.entries(arg0);
+    return ret;
+}
+export function __wbg_get_3e9a707ab7d352eb() { return handleError(function (arg0, arg1) {
+    const ret = Reflect.get(arg0, arg1);
+    return ret;
+}, arguments); }
+export function __wbg_get_98fdf51d029a75eb(arg0, arg1) {
+    const ret = arg0[arg1 >>> 0];
+    return ret;
+}
+export function __wbg_get_unchecked_1dfe6d05ad91d9b7(arg0, arg1) {
+    const ret = arg0[arg1 >>> 0];
+    return ret;
+}
+export function __wbg_get_with_ref_key_6412cf3094599694(arg0, arg1) {
+    const ret = arg0[arg1];
+    return ret;
+}
+export function __wbg_instanceof_ArrayBuffer_53db37b06f6b9afe(arg0) {
+    let result;
+    try {
+        result = arg0 instanceof ArrayBuffer;
+    } catch (_) {
+        result = false;
+    }
+    const ret = result;
+    return ret;
+}
+export function __wbg_instanceof_Map_16f217b9a2a08d8c(arg0) {
+    let result;
+    try {
+        result = arg0 instanceof Map;
+    } catch (_) {
+        result = false;
+    }
+    const ret = result;
+    return ret;
+}
+export function __wbg_instanceof_Uint8Array_abd07d4bd221d50b(arg0) {
+    let result;
+    try {
+        result = arg0 instanceof Uint8Array;
+    } catch (_) {
+        result = false;
+    }
+    const ret = result;
+    return ret;
+}
+export function __wbg_isArray_94898ed3aad6947b(arg0) {
+    const ret = Array.isArray(arg0);
+    return ret;
+}
+export function __wbg_isSafeInteger_01e964d144ad3a55(arg0) {
+    const ret = Number.isSafeInteger(arg0);
+    return ret;
+}
+export function __wbg_iterator_1441b47f341dc34f() {
+    const ret = Symbol.iterator;
+    return ret;
+}
+export function __wbg_length_2591a0f4f659a55c(arg0) {
+    const ret = arg0.length;
+    return ret;
+}
+export function __wbg_length_56fcd3e2b7e0299d(arg0) {
+    const ret = arg0.length;
+    return ret;
+}
+export function __wbg_new_02d162bc6cf02f60() {
+    const ret = new Object();
+    return ret;
+}
+export function __wbg_new_310879b66b6e95e1() {
+    const ret = new Array();
+    return ret;
+}
+export function __wbg_new_7ddec6de44ff8f5d(arg0) {
+    const ret = new Uint8Array(arg0);
+    return ret;
+}
+export function __wbg_next_2a4e19f4f5083b0f(arg0) {
+    const ret = arg0.next;
+    return ret;
+}
+export function __wbg_next_6429a146bf756f93() { return handleError(function (arg0) {
+    const ret = arg0.next();
+    return ret;
+}, arguments); }
+export function __wbg_prototypesetcall_5f9bdc8d75e07276(arg0, arg1, arg2) {
+    Uint8Array.prototype.set.call(getArrayU8FromWasm0(arg0, arg1), arg2);
+}
+export function __wbg_set_6be42768c690e380(arg0, arg1, arg2) {
+    arg0[arg1] = arg2;
+}
+export function __wbg_set_78ea6a19f4818587(arg0, arg1, arg2) {
+    arg0[arg1 >>> 0] = arg2;
+}
+export function __wbg_value_9cc0518af87a489c(arg0) {
+    const ret = arg0.value;
+    return ret;
+}
+export function __wbindgen_cast_0000000000000001(arg0) {
+    // Cast intrinsic for `F64 -> Externref`.
+    const ret = arg0;
+    return ret;
+}
+export function __wbindgen_cast_0000000000000002(arg0) {
+    // Cast intrinsic for `I64 -> Externref`.
+    const ret = arg0;
+    return ret;
+}
+export function __wbindgen_cast_0000000000000003(arg0, arg1) {
+    // Cast intrinsic for `Ref(String) -> Externref`.
+    const ret = getStringFromWasm0(arg0, arg1);
+    return ret;
+}
+export function __wbindgen_cast_0000000000000004(arg0) {
+    // Cast intrinsic for `U64 -> Externref`.
+    const ret = BigInt.asUintN(64, arg0);
+    return ret;
+}
+export function __wbindgen_init_externref_table() {
+    const table = wasm.__wbindgen_externrefs;
+    const offset = table.grow(4);
+    table.set(0, undefined);
+    table.set(offset + 0, undefined);
+    table.set(offset + 1, null);
+    table.set(offset + 2, true);
+    table.set(offset + 3, false);
+}
+const QueryMatcherFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_querymatcher_free(ptr, 1));
+
+function addToExternrefTable0(obj) {
+    const idx = wasm.__externref_table_alloc();
+    wasm.__wbindgen_externrefs.set(idx, obj);
+    return idx;
+}
+
+function debugString(val) {
+    // primitive types
+    const type = typeof val;
+    if (type == 'number' || type == 'boolean' || val == null) {
+        return  `${val}`;
+    }
+    if (type == 'string') {
+        return `"${val}"`;
+    }
+    if (type == 'symbol') {
+        const description = val.description;
+        if (description == null) {
+            return 'Symbol';
+        } else {
+            return `Symbol(${description})`;
+        }
+    }
+    if (type == 'function') {
+        const name = val.name;
+        if (typeof name == 'string' && name.length > 0) {
+            return `Function(${name})`;
+        } else {
+            return 'Function';
+        }
+    }
+    // objects
+    if (Array.isArray(val)) {
+        const length = val.length;
+        let debug = '[';
+        if (length > 0) {
+            debug += debugString(val[0]);
+        }
+        for(let i = 1; i < length; i++) {
+            debug += ', ' + debugString(val[i]);
+        }
+        debug += ']';
+        return debug;
+    }
+    // Test for built-in
+    const builtInMatches = /\[object ([^\]]+)\]/.exec(toString.call(val));
+    let className;
+    if (builtInMatches && builtInMatches.length > 1) {
+        className = builtInMatches[1];
+    } else {
+        // Failed to match the standard '[object ClassName]'
+        return toString.call(val);
+    }
+    if (className == 'Object') {
+        // we're a user defined class or Object
+        // JSON.stringify avoids problems with cycles, and is generally much
+        // easier than looping through ownProperties of `val`.
+        try {
+            return 'Object(' + JSON.stringify(val) + ')';
+        } catch (_) {
+            return 'Object';
+        }
+    }
+    // errors
+    if (val instanceof Error) {
+        return `${val.name}: ${val.message}\n${val.stack}`;
+    }
+    // TODO we could test for more things here, like `Set`s and `Map`s.
+    return className;
+}
+
+function getArrayU32FromWasm0(ptr, len) {
+    ptr = ptr >>> 0;
+    return getUint32ArrayMemory0().subarray(ptr / 4, ptr / 4 + len);
+}
+
+function getArrayU8FromWasm0(ptr, len) {
+    ptr = ptr >>> 0;
+    return getUint8ArrayMemory0().subarray(ptr / 1, ptr / 1 + len);
+}
+
+let cachedDataViewMemory0 = null;
+function getDataViewMemory0() {
+    if (cachedDataViewMemory0 === null || cachedDataViewMemory0.buffer.detached === true || (cachedDataViewMemory0.buffer.detached === undefined && cachedDataViewMemory0.buffer !== wasm.memory.buffer)) {
+        cachedDataViewMemory0 = new DataView(wasm.memory.buffer);
+    }
+    return cachedDataViewMemory0;
+}
+
+function getStringFromWasm0(ptr, len) {
+    return decodeText(ptr >>> 0, len);
+}
+
+let cachedUint32ArrayMemory0 = null;
+function getUint32ArrayMemory0() {
+    if (cachedUint32ArrayMemory0 === null || cachedUint32ArrayMemory0.byteLength === 0) {
+        cachedUint32ArrayMemory0 = new Uint32Array(wasm.memory.buffer);
+    }
+    return cachedUint32ArrayMemory0;
+}
+
+let cachedUint8ArrayMemory0 = null;
+function getUint8ArrayMemory0() {
+    if (cachedUint8ArrayMemory0 === null || cachedUint8ArrayMemory0.byteLength === 0) {
+        cachedUint8ArrayMemory0 = new Uint8Array(wasm.memory.buffer);
+    }
+    return cachedUint8ArrayMemory0;
+}
+
+function handleError(f, args) {
+    try {
+        return f.apply(this, args);
+    } catch (e) {
+        const idx = addToExternrefTable0(e);
+        wasm.__wbindgen_exn_store(idx);
+    }
+}
+
+function isLikeNone(x) {
+    return x === undefined || x === null;
+}
+
+function passArray32ToWasm0(arg, malloc) {
+    const ptr = malloc(arg.length * 4, 4) >>> 0;
+    getUint32ArrayMemory0().set(arg, ptr / 4);
+    WASM_VECTOR_LEN = arg.length;
+    return ptr;
+}
+
+function passArrayJsValueToWasm0(array, malloc) {
+    const ptr = malloc(array.length * 4, 4) >>> 0;
+    for (let i = 0; i < array.length; i++) {
+        const add = addToExternrefTable0(array[i]);
+        getDataViewMemory0().setUint32(ptr + 4 * i, add, true);
+    }
+    WASM_VECTOR_LEN = array.length;
+    return ptr;
+}
+
+function passStringToWasm0(arg, malloc, realloc) {
+    if (realloc === undefined) {
+        const buf = cachedTextEncoder.encode(arg);
+        const ptr = malloc(buf.length, 1) >>> 0;
+        getUint8ArrayMemory0().subarray(ptr, ptr + buf.length).set(buf);
+        WASM_VECTOR_LEN = buf.length;
+        return ptr;
+    }
+
+    let len = arg.length;
+    let ptr = malloc(len, 1) >>> 0;
+
+    const mem = getUint8ArrayMemory0();
+
+    let offset = 0;
+
+    for (; offset < len; offset++) {
+        const code = arg.charCodeAt(offset);
+        if (code > 0x7F) break;
+        mem[ptr + offset] = code;
+    }
+    if (offset !== len) {
+        if (offset !== 0) {
+            arg = arg.slice(offset);
+        }
+        ptr = realloc(ptr, len, len = offset + arg.length * 3, 1) >>> 0;
+        const view = getUint8ArrayMemory0().subarray(ptr + offset, ptr + len);
+        const ret = cachedTextEncoder.encodeInto(arg, view);
+
+        offset += ret.written;
+        ptr = realloc(ptr, len, offset, 1) >>> 0;
+    }
+
+    WASM_VECTOR_LEN = offset;
+    return ptr;
+}
+
+function takeFromExternrefTable0(idx) {
+    const value = wasm.__wbindgen_externrefs.get(idx);
+    wasm.__externref_table_dealloc(idx);
+    return value;
+}
+
+let cachedTextDecoder = new TextDecoder('utf-8', { ignoreBOM: true, fatal: true });
+cachedTextDecoder.decode();
+const MAX_SAFARI_DECODE_BYTES = 2146435072;
+let numBytesDecoded = 0;
+function decodeText(ptr, len) {
+    numBytesDecoded += len;
+    if (numBytesDecoded >= MAX_SAFARI_DECODE_BYTES) {
+        cachedTextDecoder = new TextDecoder('utf-8', { ignoreBOM: true, fatal: true });
+        cachedTextDecoder.decode();
+        numBytesDecoded = len;
+    }
+    return cachedTextDecoder.decode(getUint8ArrayMemory0().subarray(ptr, ptr + len));
+}
+
+const cachedTextEncoder = new TextEncoder();
+
+if (!('encodeInto' in cachedTextEncoder)) {
+    cachedTextEncoder.encodeInto = function (arg, view) {
+        const buf = cachedTextEncoder.encode(arg);
+        view.set(buf);
+        return {
+            read: arg.length,
+            written: buf.length
+        };
+    };
+}
+
+let WASM_VECTOR_LEN = 0;
+
+
+let wasm;
+export function __wbg_set_wasm(val) {
+    wasm = val;
+}
diff --git a/crates/yaak-query/pkg/yaak_query_bg.wasm b/crates/yaak-query/pkg/yaak_query_bg.wasm
new file mode 100644
index 000000000..c71e7d627
Binary files /dev/null and b/crates/yaak-query/pkg/yaak_query_bg.wasm differ
diff --git a/crates/yaak-query/pkg/yaak_query_bg.wasm.d.ts b/crates/yaak-query/pkg/yaak_query_bg.wasm.d.ts
new file mode 100644
index 000000000..ad8a62c71
--- /dev/null
+++ b/crates/yaak-query/pkg/yaak_query_bg.wasm.d.ts
@@ -0,0 +1,21 @@
+/* tslint:disable */
+/* eslint-disable */
+export const memory: WebAssembly.Memory;
+export const __wbg_querymatcher_free: (a: number, b: number) => void;
+export const parse_query: (a: number, b: number) => [number, number, number];
+export const query_completions: (a: number, b: number, c: number, d: any, e: any) => [number, number, number];
+export const query_diagnostics: (a: number, b: number, c: any) => [number, number, number];
+export const query_format_field: (a: number, b: number, c: number, d: number, e: number) => [number, number];
+export const query_tokens: (a: number, b: number) => [number, number, number];
+export const querymatcher_isEmpty: (a: number) => number;
+export const querymatcher_matches: (a: number, b: number, c: number) => [number, number, number];
+export const querymatcher_matchingRows: (a: number, b: number, c: number, d: number, e: number, f: number, g: number) => [number, number, number, number];
+export const querymatcher_new: (a: number, b: number, c: any) => [number, number, number];
+export const __wbindgen_malloc: (a: number, b: number) => number;
+export const __wbindgen_realloc: (a: number, b: number, c: number, d: number) => number;
+export const __wbindgen_exn_store: (a: number) => void;
+export const __externref_table_alloc: () => number;
+export const __wbindgen_externrefs: WebAssembly.Table;
+export const __externref_table_dealloc: (a: number) => void;
+export const __wbindgen_free: (a: number, b: number, c: number) => void;
+export const __wbindgen_start: () => void;
diff --git a/crates/yaak-query/src/ast.rs b/crates/yaak-query/src/ast.rs
new file mode 100644
index 000000000..0b1164154
--- /dev/null
+++ b/crates/yaak-query/src/ast.rs
@@ -0,0 +1,195 @@
+//! The parsed query tree.
+//!
+//! Every node carries the byte span of its source text, so a consumer can highlight it, point
+//! at it in a diagnostic, or rewrite it in place. A consumer that compiles queries (to SQL, say)
+//! walks [`Node`] directly; nothing here depends on how the query is evaluated.
+
+use serde::{Deserialize, Serialize};
+
+/// A half-open byte range `start..end` in the query source.
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub struct Span {
+    pub start: u32,
+    pub end: u32,
+}
+
+impl Span {
+    pub fn new(start: usize, end: usize) -> Self {
+        Self { start: start as u32, end: end as u32 }
+    }
+
+    pub fn start(&self) -> usize {
+        self.start as usize
+    }
+
+    pub fn end(&self) -> usize {
+        self.end as usize
+    }
+
+    /// The source text this span covers.
+    pub fn slice<'a>(&self, source: &'a str) -> &'a str {
+        &source[self.start()..self.end()]
+    }
+
+    pub(crate) fn to(&self, other: Span) -> Span {
+        Span { start: self.start.min(other.start), end: self.end.max(other.end) }
+    }
+}
+
+/// A node of the parsed query.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[serde(tag = "type")]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum Node {
+    /// Every child matches: terms side by side, or joined with `AND`.
+    And { children: Vec<Node>, span: Span },
+    /// Any child matches: terms joined with `OR`.
+    Or { children: Vec<Node>, span: Span },
+    /// The node does not match: `-term`, `-(group)`, `-@field:value`.
+    Not { node: Box<Node>, span: Span },
+    /// A bare word, matched against the document's free text.
+    Term { text: String, span: Span },
+    /// A double-quoted string, matched against the free text with its spaces kept.
+    Phrase { text: String, span: Span },
+    /// `@path:value`.
+    Field {
+        /// The dotted path, one entry per identifier: `@req.header.accept` is
+        /// `["req", "header", "accept"]`.
+        path: Vec<String>,
+        value: FieldValue,
+        span: Span,
+        /// `@path:`, from the `@` through the colon.
+        #[serde(rename = "pathSpan")]
+        path_span: Span,
+        /// The value as written, quotes included. Empty (and at the end of the field) when
+        /// nothing follows the colon.
+        #[serde(rename = "valueSpan")]
+        value_span: Span,
+    },
+}
+
+impl Node {
+    pub fn span(&self) -> Span {
+        match self {
+            Node::And { span, .. }
+            | Node::Or { span, .. }
+            | Node::Not { span, .. }
+            | Node::Term { span, .. }
+            | Node::Phrase { span, .. }
+            | Node::Field { span, .. } => *span,
+        }
+    }
+
+    /// Calls `f` on this node and every node below it, parents first.
+    pub fn walk<'a>(&'a self, f: &mut impl FnMut(&'a Node)) {
+        f(self);
+        match self {
+            Node::And { children, .. } | Node::Or { children, .. } => {
+                for child in children {
+                    child.walk(f);
+                }
+            }
+            Node::Not { node, .. } => node.walk(f),
+            Node::Term { .. } | Node::Phrase { .. } | Node::Field { .. } => {}
+        }
+    }
+
+    /// Rewrites every span in the tree, for a consumer that counts positions differently (the
+    /// WebAssembly package hands JavaScript UTF-16 offsets).
+    pub fn map_spans(&mut self, f: &impl Fn(Span) -> Span) {
+        match self {
+            Node::And { children, span } | Node::Or { children, span } => {
+                *span = f(*span);
+                for child in children {
+                    child.map_spans(f);
+                }
+            }
+            Node::Not { node, span } => {
+                *span = f(*span);
+                node.map_spans(f);
+            }
+            Node::Term { span, .. } | Node::Phrase { span, .. } => *span = f(*span),
+            Node::Field { span, path_span, value_span, .. } => {
+                *span = f(*span);
+                *path_span = f(*path_span);
+                *value_span = f(*value_span);
+            }
+        }
+    }
+}
+
+/// What a field term's value asks for. The shape is decided by the parser alone; the schema
+/// decides what it means for a given field.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "camelCase")]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum FieldValue {
+    /// Plain text, matched as a case-insensitive substring. `quoted` when it was written in
+    /// double quotes (escapes already removed). Empty for a trailing `@field:`.
+    Text { text: String, quoted: bool },
+    /// A bare value containing `*`, which must match the whole field value. `*` matches any
+    /// run of characters, including none.
+    Wildcard { pattern: String },
+    /// `*` alone: the field has a value. Negated, it has none.
+    Exists,
+    /// `>`, `>=`, `<` or `<=` followed by a number.
+    Compare {
+        op: CompareOp,
+        number: NumberLiteral,
+    },
+    /// `[low TO high]`, inclusive.
+    Range {
+        low: NumberLiteral,
+        high: NumberLiteral,
+    },
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum CompareOp {
+    #[serde(rename = ">")]
+    Gt,
+    #[serde(rename = ">=")]
+    Gte,
+    #[serde(rename = "<")]
+    Lt,
+    #[serde(rename = "<=")]
+    Lte,
+}
+
+impl CompareOp {
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            CompareOp::Gt => ">",
+            CompareOp::Gte => ">=",
+            CompareOp::Lt => "<",
+            CompareOp::Lte => "<=",
+        }
+    }
+
+    pub fn test(&self, left: f64, right: f64) -> bool {
+        match self {
+            CompareOp::Gt => left > right,
+            CompareOp::Gte => left >= right,
+            CompareOp::Lt => left < right,
+            CompareOp::Lte => left <= right,
+        }
+    }
+}
+
+/// A number as written in a comparison or range, with its unit when it has one (`500ms`,
+/// `2.5mb`). The schema says which units a field accepts and what they are worth.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub struct NumberLiteral {
+    pub value: f64,
+    pub unit: Option<String>,
+    /// The number as it appears in the source, unit included.
+    pub raw: String,
+}
diff --git a/crates/yaak-query/src/complete.rs b/crates/yaak-query/src/complete.rs
new file mode 100644
index 000000000..9849c50c2
--- /dev/null
+++ b/crates/yaak-query/src/complete.rs
@@ -0,0 +1,378 @@
+//! What can complete the term at the cursor.
+
+use crate::ast::Span;
+use crate::format::{format_value, is_valid_path};
+use crate::lexer::{TokKind, is_ident_char, lex, lex_quoted};
+use crate::schema::{FieldKind, Schema};
+use serde::{Deserialize, Serialize};
+use std::collections::{HashMap, HashSet};
+
+/// Completions for the term at the cursor, all replacing `span`.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub struct Completions {
+    pub span: Span,
+    pub items: Vec<Completion>,
+}
+
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub struct Completion {
+    /// What to show: a field's path, or a value as it is.
+    pub label: String,
+    /// What replaces the span: `@method:`, `@header.`, or a value quoted when it needs to be.
+    pub insert: String,
+    pub kind: CompletionKind,
+    pub info: Option<String>,
+    /// The term is not finished after inserting (a field still needs its value, a map field its
+    /// key), so completion should open again.
+    pub reopen: bool,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum CompletionKind {
+    Field,
+    Value,
+}
+
+/// Completions at byte offset `cursor` of `input`.
+///
+/// At the start of a term, or in a word made of identifier characters (with or without its
+/// `@`), these are the schema's fields whose path contains what is typed, those starting with
+/// it first. After `@map.` they are the keys given in `values` under the map's name. After
+/// `@field:` they are the values given in `values` under the field's path that contain what is
+/// typed (quotes and escapes aside); the whole value is replaced. A field written without its
+/// `@` (`method:post`) gets the term with the `@`. Inside a phrase, after a `)`, or in other
+/// text there are none.
+pub fn completions(
+    input: &str,
+    cursor: usize,
+    schema: &Schema,
+    values: &HashMap<String, Vec<String>>,
+) -> Option<Completions> {
+    let mut cursor = cursor.min(input.len());
+    while !input.is_char_boundary(cursor) {
+        cursor -= 1;
+    }
+
+    let toks = lex(input);
+    let Some(tok) = toks.iter().find(|t| t.span.start() < cursor && cursor <= t.span.end()) else {
+        // Between terms: anything may start here
+        let at_term_start = input[..cursor].chars().next_back().is_none_or(char::is_whitespace);
+        return at_term_start
+            .then(|| field_completions(schema, values, "", Span::new(cursor, cursor)))
+            .flatten();
+    };
+
+    match &tok.kind {
+        TokKind::Minus | TokKind::LParen if cursor == tok.span.end() => {
+            field_completions(schema, values, "", Span::new(cursor, cursor))
+        }
+        TokKind::Word(word) => {
+            let typed = &input[tok.span.start()..cursor];
+            if let Some(fix) = missing_at(word, typed, schema) {
+                return Some(Completions { span: tok.span, items: vec![fix] });
+            }
+            let partial = typed.strip_prefix('@').unwrap_or(typed);
+            if !partial.chars().all(|c| is_ident_char(c) || c == '.') {
+                return None;
+            }
+            field_completions(schema, values, partial, Span::new(tok.span.start(), cursor))
+        }
+        TokKind::Field { path, path_span, value_span, .. } => {
+            if cursor < value_span.start() {
+                // In the path: replace it through the colon
+                let partial = &input[tok.span.start() + 1..cursor];
+                return field_completions(schema, values, partial, *path_span);
+            }
+            value_completions(input, cursor, path, *value_span, schema, values)
+        }
+        _ => None,
+    }
+}
+
+/// A word written like a field but without its `@` (`method:post`), once the cursor is past the
+/// colon: the same term with the `@`, replacing the word. Words like `localhost:3000`, whose
+/// part before the colon is no field, get nothing.
+fn missing_at(word: &str, typed: &str, schema: &Schema) -> Option<Completion> {
+    if word.starts_with('@') || !typed.contains(':') {
+        return None;
+    }
+    let (path, value) = word.split_once(':')?;
+    if !is_valid_path(path) {
+        return None;
+    }
+    let path: Vec<String> = path.split('.').map(String::from).collect();
+    schema.resolve(&path)?;
+    let term = format!("@{word}");
+    Some(Completion {
+        label: term.clone(),
+        insert: term,
+        kind: CompletionKind::Field,
+        info: None,
+        // With no value yet, go on to the field's values
+        reopen: value.is_empty(),
+    })
+}
+
+fn field_completions(
+    schema: &Schema,
+    values: &HashMap<String, Vec<String>>,
+    partial: &str,
+    span: Span,
+) -> Option<Completions> {
+    let needle = partial.to_lowercase();
+    let mut items = Vec::new();
+
+    // Keys under a map: `@header.x-` offers `@header.x-cache:`
+    if let Some((map_name, key_partial)) = needle.rsplit_once('.')
+        && let Some(map) = schema.fields.iter().find(|f| f.map && f.name.to_lowercase() == map_name)
+    {
+        let keys = lookup(values, map_name).into_iter().flatten();
+        for key in ranked(keys, key_partial) {
+            let path = format!("{}.{}", map.name, key);
+            if !is_valid_path(&path) {
+                continue;
+            }
+            items.push(Completion {
+                insert: format!("@{path}:"),
+                label: path,
+                kind: CompletionKind::Field,
+                info: None,
+                reopen: true,
+            });
+        }
+    }
+
+    for field in ranked_by(&schema.fields, &needle, |f| f.name.as_str()) {
+        let insert =
+            if field.map { format!("@{}.", field.name) } else { format!("@{}:", field.name) };
+        items.push(Completion {
+            label: field.name.clone(),
+            insert,
+            kind: CompletionKind::Field,
+            info: field.info.clone(),
+            reopen: true,
+        });
+    }
+
+    (!items.is_empty()).then_some(Completions { span, items })
+}
+
+fn value_completions(
+    input: &str,
+    cursor: usize,
+    path: &[String],
+    value_span: Span,
+    schema: &Schema,
+    values: &HashMap<String, Vec<String>>,
+) -> Option<Completions> {
+    let field = schema.resolve(path)?;
+    let typed = &input[value_span.start()..cursor];
+    let needle = if typed.starts_with('"') { lex_quoted(typed, 0).0 } else { typed.to_string() };
+
+    let joined = path.join(".");
+    let flags = ["true".to_string(), "false".to_string()];
+    let candidates: Vec<&String> = match lookup(values, &joined) {
+        Some(given) => given.iter().collect(),
+        None if field.def.kind == FieldKind::Flag => flags.iter().collect(),
+        None => return None,
+    };
+
+    let items: Vec<Completion> = ranked(candidates, &needle.to_lowercase())
+        .into_iter()
+        .map(|value| Completion {
+            label: value.clone(),
+            insert: format_value(value),
+            kind: CompletionKind::Value,
+            info: None,
+            reopen: false,
+        })
+        .collect();
+    (!items.is_empty()).then_some(Completions { span: value_span, items })
+}
+
+/// The values a consumer gave for `path`, matched case-insensitively.
+fn lookup<'a>(values: &'a HashMap<String, Vec<String>>, path: &str) -> Option<&'a Vec<String>> {
+    values
+        .get(path)
+        .or_else(|| values.iter().find(|(k, _)| k.eq_ignore_ascii_case(path)).map(|(_, v)| v))
+}
+
+/// The strings containing `needle` (lowercase), those starting with it first, each once.
+fn ranked<'a>(items: impl IntoIterator<Item = &'a String>, needle: &str) -> Vec<&'a String> {
+    let items: Vec<&String> = items.into_iter().collect();
+    let mut seen = HashSet::new();
+    ranked_by(&items, needle, |s| s.as_str())
+        .into_iter()
+        .copied()
+        .filter(|s| seen.insert(*s))
+        .collect()
+}
+
+fn ranked_by<'a, T>(items: &'a [T], needle: &str, key: impl Fn(&T) -> &str) -> Vec<&'a T> {
+    let lower = |item: &T| key(item).to_lowercase();
+    let starts = items.iter().filter(|i| lower(i).starts_with(needle));
+    let contains = items.iter().filter(|i| {
+        let l = lower(i);
+        !l.starts_with(needle) && l.contains(needle)
+    });
+    starts.chain(contains).collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::schema::{FieldDef, FreeText};
+
+    fn schema() -> Schema {
+        let text = |name: &str| FieldDef {
+            name: name.into(),
+            kind: FieldKind::Text,
+            map: false,
+            info: None,
+        };
+        Schema::new(
+            vec![
+                text("method"),
+                text("url"),
+                text("name"),
+                FieldDef { name: "header".into(), kind: FieldKind::Text, map: true, info: None },
+                FieldDef { name: "error".into(), kind: FieldKind::Flag, map: false, info: None },
+            ],
+            FreeText::Fuzzy,
+        )
+    }
+
+    fn values() -> HashMap<String, Vec<String>> {
+        HashMap::from([
+            (
+                "method".to_string(),
+                vec!["GET".to_string(), "POST".to_string(), "PATCH".to_string()],
+            ),
+            ("name".to_string(), vec!["My Request".to_string(), "Other".to_string()]),
+            (
+                "header".to_string(),
+                vec![
+                    "x-cache".to_string(),
+                    "Accept".to_string(),
+                    "bad key".to_string(),
+                ],
+            ),
+        ])
+    }
+
+    /// The completion at the `|` in `input`: the replaced text and the inserts.
+    fn at(input: &str) -> Option<(String, Vec<String>)> {
+        let cursor = input.find('|').unwrap();
+        let input = input.replacen('|', "", 1);
+        let c = completions(&input, cursor, &schema(), &values())?;
+        Some((c.span.slice(&input).to_string(), c.items.into_iter().map(|i| i.insert).collect()))
+    }
+
+    fn inserts(list: &[&str]) -> Vec<String> {
+        list.iter().map(|s| s.to_string()).collect()
+    }
+
+    #[test]
+    fn fields_at_a_term_start() {
+        let all = inserts(&["@method:", "@url:", "@name:", "@header.", "@error:"]);
+        assert_eq!(at("|"), Some(("".into(), all.clone())));
+        assert_eq!(at("a |"), Some(("".into(), all.clone())));
+        assert_eq!(at("-|"), Some(("".into(), all.clone())));
+        assert_eq!(at("(|"), Some(("".into(), all)));
+    }
+
+    #[test]
+    fn fields_for_a_partial_word() {
+        assert_eq!(at("@met|"), Some(("@met".into(), inserts(&["@method:"]))));
+        assert_eq!(at("met|"), Some(("met".into(), inserts(&["@method:"]))));
+        // Starts-with first, then contains
+        assert_eq!(at("@me|"), Some(("@me".into(), inserts(&["@method:", "@name:"]))));
+        assert_eq!(at("-@ur|"), Some(("@ur".into(), inserts(&["@url:"]))));
+        assert_eq!(at("x (@na|"), Some(("@na".into(), inserts(&["@name:"]))));
+        assert_eq!(
+            at("@e|"),
+            Some(("@e".into(), inserts(&["@error:", "@method:", "@name:", "@header."])))
+        );
+        assert_eq!(at("@zz|"), None);
+        assert_eq!(at("localhost:30|"), None);
+    }
+
+    #[test]
+    fn map_keys() {
+        assert_eq!(
+            at("@header.|"),
+            Some(("@header.".into(), inserts(&["@header.x-cache:", "@header.Accept:"])))
+        );
+        assert_eq!(
+            at("@header.ac|"),
+            Some(("@header.ac".into(), inserts(&["@header.Accept:", "@header.x-cache:"])))
+        );
+    }
+
+    #[test]
+    fn values_after_the_colon() {
+        assert_eq!(at("@method:|"), Some(("".into(), inserts(&["GET", "POST", "PATCH"]))));
+        assert_eq!(at("@method:p|"), Some(("p".into(), inserts(&["POST", "PATCH"]))));
+        assert_eq!(at("@method:t|"), Some(("t".into(), inserts(&["GET", "POST", "PATCH"]))));
+        assert_eq!(at("@name:my|"), Some(("my".into(), inserts(&["\"My Request\""]))));
+        assert_eq!(at("@name:\"my r|"), Some(("\"my r".into(), inserts(&["\"My Request\""]))));
+        // The whole value is replaced, even past the cursor
+        assert_eq!(at("@method:p|os x"), Some(("pos".into(), inserts(&["POST", "PATCH"]))));
+        assert_eq!(at("@method:zz|"), None);
+        // Fields with no values given offer none, unknown fields neither
+        assert_eq!(at("@url:|"), None);
+        assert_eq!(at("@nope:|"), None);
+        // Flags offer true and false
+        assert_eq!(at("@error:|"), Some(("".into(), inserts(&["true", "false"]))));
+    }
+
+    #[test]
+    fn editing_a_field_name() {
+        assert_eq!(at("@met|hod:post"), Some(("@method:".into(), inserts(&["@method:"]))));
+    }
+
+    #[test]
+    fn suggests_the_at_for_a_field_written_without_it() {
+        assert_eq!(at("method:post|"), Some(("method:post".into(), inserts(&["@method:post"]))));
+        assert_eq!(at("x method:p|"), Some(("method:p".into(), inserts(&["@method:p"]))));
+        assert_eq!(at("-method:po|st"), Some(("method:post".into(), inserts(&["@method:post"]))));
+        assert_eq!(at("(METHOD:get|"), Some(("METHOD:get".into(), inserts(&["@METHOD:get"]))));
+        assert_eq!(
+            at("header.accept:json|"),
+            Some(("header.accept:json".into(), inserts(&["@header.accept:json"])))
+        );
+        // With nothing after the colon, the field's values come next
+        let c = completions("method:", 7, &schema(), &values()).unwrap();
+        assert_eq!(c.items[0].insert, "@method:");
+        assert!(c.items[0].reopen);
+        assert!(!completions("method:x", 8, &schema(), &values()).unwrap().items[0].reopen);
+        // Not a field: plain text, nothing to suggest
+        assert_eq!(at("localhost:3000|"), None);
+        assert_eq!(at("nope:x|"), None);
+        assert_eq!(at("http://x|"), None);
+        // Before the colon, the word still completes field names
+        assert_eq!(at("met|hod:post"), Some(("met".into(), inserts(&["@method:"]))));
+    }
+
+    #[test]
+    fn nothing_inside_phrases_or_after_parens() {
+        assert_eq!(at("\"my |"), None);
+        assert_eq!(at("(a)|"), None);
+        // A word is replaced up to the cursor
+        assert_eq!(at("a|b"), Some(("a".into(), inserts(&["@name:", "@header."]))));
+    }
+
+    #[test]
+    fn cursor_is_clamped() {
+        let c = completions("é", 1, &schema(), &values());
+        assert!(c.is_none() || c.unwrap().span.end <= 2);
+        assert!(completions("", 99, &schema(), &values()).is_some());
+    }
+}
diff --git a/crates/yaak-query/src/eval.rs b/crates/yaak-query/src/eval.rs
new file mode 100644
index 000000000..85dd998e8
--- /dev/null
+++ b/crates/yaak-query/src/eval.rs
@@ -0,0 +1,1055 @@
+//! Runs a parsed query in memory, against one document or a whole table of them.
+
+use crate::ast::{FieldValue, Node};
+use crate::fuzzy::{FuzzyScratch, fuzzy_match_in};
+use crate::schema::{FieldKind, FreeText, Schema, parse_flag, to_base};
+use serde::{Deserialize, Serialize};
+use std::borrow::Cow;
+use std::collections::HashMap;
+
+/// What a query is matched against: free text for bare words and phrases, and field values by
+/// lowercase dotted path. A map field's values sit under `name.key` (`header.x-cache`).
+#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub struct Document {
+    #[serde(default)]
+    #[cfg_attr(test, ts(optional, as = "Option<String>"))]
+    pub text: String,
+    #[serde(default)]
+    #[cfg_attr(test, ts(optional, as = "Option<HashMap<String, DocValues>>"))]
+    pub fields: HashMap<String, DocValues>,
+}
+
+/// One value or several.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[serde(untagged)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum DocValues {
+    One(DocValue),
+    Many(Vec<DocValue>),
+}
+
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[serde(untagged)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum DocValue {
+    Flag(bool),
+    Number(f64),
+    Text(String),
+}
+
+impl Document {
+    pub fn new(text: impl Into<String>) -> Self {
+        Self { text: text.into(), fields: HashMap::new() }
+    }
+
+    /// Adds a value under `path`, keeping any already there.
+    pub fn with(mut self, path: &str, value: impl Into<DocValue>) -> Self {
+        let value = value.into();
+        let path = path.to_lowercase();
+        match self.fields.remove(&path) {
+            None => self.fields.insert(path, DocValues::One(value)),
+            Some(DocValues::One(first)) => {
+                self.fields.insert(path, DocValues::Many(vec![first, value]))
+            }
+            Some(DocValues::Many(mut values)) => {
+                values.push(value);
+                self.fields.insert(path, DocValues::Many(values))
+            }
+        };
+        self
+    }
+}
+
+impl From<&str> for DocValue {
+    fn from(value: &str) -> Self {
+        DocValue::Text(value.to_string())
+    }
+}
+
+impl From<String> for DocValue {
+    fn from(value: String) -> Self {
+        DocValue::Text(value)
+    }
+}
+
+impl From<f64> for DocValue {
+    fn from(value: f64) -> Self {
+        DocValue::Number(value)
+    }
+}
+
+impl From<bool> for DocValue {
+    fn from(value: bool) -> Self {
+        DocValue::Flag(value)
+    }
+}
+
+impl DocValue {
+    fn view(&self) -> Val<'_> {
+        match self {
+            DocValue::Text(text) => Val::Text(text),
+            DocValue::Number(n) => Val::Number(*n),
+            DocValue::Flag(b) => Val::Flag(*b),
+        }
+    }
+}
+
+/// Many documents with the same fields, as one flat run of cells: for each row, its free text,
+/// then a text value (or none) per column. Cells may borrow from the input they were read from.
+/// Matching a table is much cheaper than matching its rows one document at a time.
+#[derive(Clone, Debug, Default, PartialEq)]
+pub struct Table<'a> {
+    /// Lowercase field paths, one per column.
+    columns: Vec<String>,
+    cells: Vec<Option<Cow<'a, str>>>,
+}
+
+impl<'a> Table<'a> {
+    /// An empty table whose columns hold the fields at these paths.
+    pub fn new(columns: impl IntoIterator<Item = impl Into<String>>) -> Self {
+        let columns = columns.into_iter().map(|c| c.into().to_lowercase()).collect();
+        Self { columns, cells: Vec::new() }
+    }
+
+    /// A table from its cells, already flat (text first, then a cell per column, row after
+    /// row). `None` when the cells don't fill whole rows.
+    pub fn from_cells(
+        columns: impl IntoIterator<Item = impl Into<String>>,
+        cells: Vec<Option<Cow<'a, str>>>,
+    ) -> Option<Self> {
+        let table = Self { cells, ..Self::new(columns) };
+        table.cells.len().is_multiple_of(table.stride()).then_some(table)
+    }
+
+    /// A table from cells packed the way JavaScript can produce them cheaply: every cell joined
+    /// into `joined`, with each cell's length in UTF-16 code units in `lengths` (or
+    /// [`Table::NO_VALUE`] for a cell with no value). Cells borrow from `joined`. `None` when
+    /// the lengths don't add up to `joined` or the cells don't fill whole rows.
+    pub fn from_packed(
+        columns: impl IntoIterator<Item = impl Into<String>>,
+        joined: &'a str,
+        lengths: &[u32],
+    ) -> Option<Self> {
+        let mut cells = Vec::with_capacity(lengths.len());
+        let ascii = joined.is_ascii();
+        let mut start = 0;
+        for &len in lengths {
+            if len == Self::NO_VALUE {
+                cells.push(None);
+                continue;
+            }
+            let end = if ascii {
+                start + len as usize
+            } else {
+                // Walk `len` UTF-16 units forward
+                let mut units = 0;
+                let mut end = start;
+                for c in joined.get(start..)?.chars() {
+                    if units >= len as usize {
+                        break;
+                    }
+                    units += c.len_utf16();
+                    end += c.len_utf8();
+                }
+                if units != len as usize {
+                    return None;
+                }
+                end
+            };
+            cells.push(Some(Cow::Borrowed(joined.get(start..end)?)));
+            start = end;
+        }
+        if start != joined.len() {
+            return None;
+        }
+        Self::from_cells(columns, cells)
+    }
+
+    /// The length [`Table::from_packed`] reads as a cell with no value.
+    pub const NO_VALUE: u32 = u32::MAX;
+
+    /// Adds a row: its free text, and a value or `None` per column, in column order. Missing
+    /// trailing cells are `None`; extra ones are dropped.
+    pub fn push(
+        &mut self,
+        text: impl Into<Cow<'a, str>>,
+        cells: impl IntoIterator<Item = Option<Cow<'a, str>>>,
+    ) {
+        self.cells.push(Some(text.into()));
+        let mut cells = cells.into_iter();
+        for _ in 0..self.columns.len() {
+            self.cells.push(cells.next().flatten());
+        }
+    }
+
+    pub fn len(&self) -> usize {
+        self.cells.len() / self.stride()
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.cells.is_empty()
+    }
+
+    fn stride(&self) -> usize {
+        self.columns.len() + 1
+    }
+}
+
+/// A field value as matching sees it, borrowed from a document or a table.
+#[derive(Clone, Copy, Debug)]
+enum Val<'a> {
+    Text(&'a str),
+    Number(f64),
+    Flag(bool),
+}
+
+impl Val<'_> {
+    /// The value as lowercase text.
+    fn lower(&self) -> Cow<'_, str> {
+        match self {
+            Val::Text(text) => Cow::Owned(text.to_lowercase()),
+            Val::Number(n) => Cow::Owned(format_number(*n)),
+            Val::Flag(b) => Cow::Borrowed(if *b { "true" } else { "false" }),
+        }
+    }
+
+    /// Whether the value, lowercased, contains `needle` (lowercase).
+    fn contains(&self, needle: &str) -> bool {
+        match self {
+            Val::Text(text) => contains_ignore_case(text, needle),
+            _ => self.lower().contains(needle),
+        }
+    }
+
+    fn as_number(&self) -> Option<f64> {
+        match self {
+            Val::Number(n) => Some(*n),
+            Val::Text(text) => text.trim().parse().ok(),
+            Val::Flag(_) => None,
+        }
+    }
+
+    fn as_flag(&self) -> Option<bool> {
+        match self {
+            Val::Flag(b) => Some(*b),
+            Val::Text(text) => parse_flag(text),
+            Val::Number(n) => Some(*n != 0.0),
+        }
+    }
+
+    fn is_present(&self) -> bool {
+        match self {
+            Val::Text(text) => !text.is_empty(),
+            Val::Number(_) => true,
+            Val::Flag(b) => *b,
+        }
+    }
+}
+
+/// Whether `hay`, lowercased, contains `needle` (lowercase), without copying ASCII text.
+fn contains_ignore_case(hay: &str, needle: &str) -> bool {
+    if !(hay.is_ascii() && needle.is_ascii()) {
+        return hay.to_lowercase().contains(needle);
+    }
+    let (hay, needle) = (hay.as_bytes(), needle.as_bytes());
+    needle.is_empty() || hay.windows(needle.len()).any(|w| w.eq_ignore_ascii_case(needle))
+}
+
+/// `text` lowercased into `out`, without a fresh allocation per call.
+fn lowercase_into(text: &str, out: &mut String) {
+    out.clear();
+    if text.is_ascii() {
+        out.push_str(text);
+        out.make_ascii_lowercase();
+    } else {
+        out.push_str(&text.to_lowercase());
+    }
+}
+
+/// `200`, not `200.0`, so a number field's text matches what the user sees.
+fn format_number(n: f64) -> String {
+    if n.fract() == 0.0 && n.abs() < 1e15 { format!("{}", n as i64) } else { n.to_string() }
+}
+
+/// Where field values come from: a [`Document`], or a row of a [`Table`].
+trait FieldSource {
+    /// Whether any value under `path` (lowercase) passes `test`. With `any_key`, `path` is a
+    /// map field and the values of every key under it count.
+    fn any_value(&self, path: &str, any_key: bool, test: &mut dyn FnMut(Val) -> bool) -> bool;
+}
+
+impl FieldSource for Document {
+    fn any_value(&self, path: &str, any_key: bool, test: &mut dyn FnMut(Val) -> bool) -> bool {
+        self.fields.iter().filter(|(key, _)| path_matches(key, path, any_key)).any(|(_, values)| {
+            match values {
+                DocValues::One(v) => test(v.view()),
+                DocValues::Many(vs) => vs.iter().any(|v| test(v.view())),
+            }
+        })
+    }
+}
+
+struct RowSource<'t, 'a> {
+    columns: &'t [String],
+    /// The row's cells, after its text.
+    cells: &'t [Option<Cow<'a, str>>],
+}
+
+impl FieldSource for RowSource<'_, '_> {
+    fn any_value(&self, path: &str, any_key: bool, test: &mut dyn FnMut(Val) -> bool) -> bool {
+        self.columns.iter().zip(self.cells).any(|(column, cell)| {
+            path_matches(column, path, any_key)
+                && cell.as_deref().is_some_and(|v| test(Val::Text(v)))
+        })
+    }
+}
+
+/// Whether a document key is `path`, or with `any_key`, any key under it (`path.<key>`). Paths
+/// are ASCII identifiers, so comparing ASCII case-insensitively is exact.
+fn path_matches(key: &str, path: &str, any_key: bool) -> bool {
+    let (key, path) = (key.as_bytes(), path.as_bytes());
+    if key.eq_ignore_ascii_case(path) {
+        return true;
+    }
+    any_key
+        && key.len() > path.len() + 1
+        && key[..path.len()].eq_ignore_ascii_case(path)
+        && key[path.len()] == b'.'
+}
+
+/// Whether `doc` matches the query. `None` (an empty query) matches everything. To match many
+/// documents, build a [`Matcher`] once instead.
+pub fn evaluate(root: Option<&Node>, schema: &Schema, doc: &Document) -> bool {
+    Matcher::new(root, schema).matches(doc)
+}
+
+/// A query resolved against a schema once (fields looked up, text lowercased), to match many
+/// documents.
+#[derive(Clone, Debug, PartialEq)]
+pub struct Matcher {
+    plan: Option<Plan>,
+    fuzzy: bool,
+    /// Whether any term reads the free text, which then has to be lowercased per document.
+    uses_text: bool,
+}
+
+#[derive(Clone, Debug, PartialEq)]
+enum Plan {
+    And(Vec<Plan>),
+    Or(Vec<Plan>),
+    Not(Box<Plan>),
+    /// A bare word, lowercased and trimmed.
+    Term(String),
+    /// A phrase, lowercased.
+    Phrase(String),
+    Field {
+        /// The lowercase path the document stores the values under.
+        path: String,
+        /// A map field with no key: every key's values.
+        any_key: bool,
+        kind: FieldKind,
+        /// With its text and wildcard pattern lowercased.
+        value: FieldValue,
+    },
+    /// A term that can match nothing: an unknown field, an empty word or phrase.
+    Never,
+}
+
+/// Buffers reused from one document to the next.
+#[derive(Default)]
+struct Scratch {
+    text: String,
+    fuzzy: FuzzyScratch,
+}
+
+impl Matcher {
+    pub fn new(root: Option<&Node>, schema: &Schema) -> Self {
+        let mut uses_text = false;
+        let plan = root.map(|root| plan_node(root, schema, &mut uses_text));
+        Self { plan, fuzzy: schema.free_text == FreeText::Fuzzy, uses_text }
+    }
+
+    /// Whether the query has no terms, so it matches everything.
+    pub fn is_empty(&self) -> bool {
+        self.plan.is_none()
+    }
+
+    pub fn matches(&self, doc: &Document) -> bool {
+        self.matches_source(&doc.text, doc, &mut Scratch::default())
+    }
+
+    /// The indexes of the rows that match, in order.
+    pub fn matching_rows(&self, table: &Table) -> Vec<u32> {
+        let mut scratch = Scratch::default();
+        let rows = table.cells.chunks_exact(table.stride());
+        rows.enumerate()
+            .filter(|(_, row)| {
+                let text = row[0].as_deref().unwrap_or_default();
+                let source = RowSource { columns: &table.columns, cells: &row[1..] };
+                self.matches_source(text, &source, &mut scratch)
+            })
+            .map(|(i, _)| i as u32)
+            .collect()
+    }
+
+    fn matches_source(&self, text: &str, source: &impl FieldSource, scratch: &mut Scratch) -> bool {
+        let Some(plan) = &self.plan else { return true };
+        if self.uses_text {
+            lowercase_into(text, &mut scratch.text);
+        }
+        let Scratch { text, fuzzy } = scratch;
+        self.eval(plan, text, source, fuzzy)
+    }
+
+    fn eval(
+        &self,
+        plan: &Plan,
+        text: &str,
+        source: &impl FieldSource,
+        fuzzy: &mut FuzzyScratch,
+    ) -> bool {
+        match plan {
+            Plan::And(children) => children.iter().all(|c| self.eval(c, text, source, fuzzy)),
+            Plan::Or(children) => children.iter().any(|c| self.eval(c, text, source, fuzzy)),
+            Plan::Not(node) => !self.eval(node, text, source, fuzzy),
+            Plan::Term(term) if self.fuzzy => fuzzy_match_in(text, term, fuzzy),
+            Plan::Term(term) | Plan::Phrase(term) => text.contains(term.as_str()),
+            Plan::Field { path, any_key, kind, value } => {
+                let mut any =
+                    |test: &mut dyn FnMut(Val) -> bool| source.any_value(path, *any_key, test);
+                match kind {
+                    FieldKind::Text => match_text(&mut any, value),
+                    FieldKind::Number { units } => match_number(&mut any, value, units),
+                    FieldKind::Flag => match_flag(&mut any, value),
+                }
+            }
+            Plan::Never => false,
+        }
+    }
+}
+
+fn plan_node(node: &Node, schema: &Schema, uses_text: &mut bool) -> Plan {
+    let mut children =
+        |nodes: &[Node]| nodes.iter().map(|n| plan_node(n, schema, uses_text)).collect();
+    match node {
+        Node::And { children: nodes, .. } => Plan::And(children(nodes)),
+        Node::Or { children: nodes, .. } => Plan::Or(children(nodes)),
+        Node::Not { node, .. } => Plan::Not(Box::new(plan_node(node, schema, uses_text))),
+        Node::Term { text, .. } => {
+            let term = text.trim().to_lowercase();
+            if term.is_empty() {
+                return Plan::Never;
+            }
+            *uses_text = true;
+            Plan::Term(term)
+        }
+        Node::Phrase { text, .. } => {
+            if text.is_empty() {
+                return Plan::Never;
+            }
+            *uses_text = true;
+            Plan::Phrase(text.to_lowercase())
+        }
+        Node::Field { path, value, .. } => {
+            let Some(field) = schema.resolve(path) else {
+                return Plan::Never;
+            };
+            let value = match value {
+                FieldValue::Text { text, quoted } => {
+                    FieldValue::Text { text: text.to_lowercase(), quoted: *quoted }
+                }
+                FieldValue::Wildcard { pattern } => {
+                    FieldValue::Wildcard { pattern: pattern.to_lowercase() }
+                }
+                other => other.clone(),
+            };
+            Plan::Field {
+                path: field.document_path(),
+                any_key: field.def.map && field.key.is_none(),
+                kind: field.def.kind.clone(),
+                value,
+            }
+        }
+    }
+}
+
+/// Asks whether any of a field's values passes a test.
+type AnyValue<'s> = dyn FnMut(&mut dyn FnMut(Val) -> bool) -> bool + 's;
+
+/// `value`'s text and wildcard pattern are already lowercase (see [`plan_node`]).
+fn match_text(any: &mut AnyValue, value: &FieldValue) -> bool {
+    match value {
+        FieldValue::Text { text: needle, .. } => {
+            !needle.is_empty() && any(&mut |v| v.contains(needle))
+        }
+        FieldValue::Wildcard { pattern } => any(&mut |v| wildcard_match(pattern, &v.lower())),
+        FieldValue::Exists => any(&mut |v| v.is_present()),
+        // A text field holding numbers, such as a status shown as text
+        FieldValue::Compare { op, number } => {
+            number.unit.is_none()
+                && any(&mut |v| v.as_number().is_some_and(|n| op.test(n, number.value)))
+        }
+        FieldValue::Range { low, high } => {
+            low.unit.is_none()
+                && high.unit.is_none()
+                && any(&mut |v| v.as_number().is_some_and(|n| n >= low.value && n <= high.value))
+        }
+    }
+}
+
+fn match_number(any: &mut AnyValue, value: &FieldValue, units: &[crate::Unit]) -> bool {
+    match value {
+        FieldValue::Text { text, quoted } => {
+            if text.is_empty() {
+                return false;
+            }
+            // `@status:404` is equality; anything else matches the number as text
+            match (!quoted).then(|| crate::lexer::parse_number(text)).flatten() {
+                Some(number) => to_base(&number, units)
+                    .is_some_and(|target| any(&mut |v| v.as_number() == Some(target))),
+                None => match_text(any, value),
+            }
+        }
+        FieldValue::Wildcard { .. } => match_text(any, value),
+        FieldValue::Exists => any(&mut |v| v.as_number().is_some()),
+        FieldValue::Compare { op, number } => to_base(number, units)
+            .is_some_and(|target| any(&mut |v| v.as_number().is_some_and(|n| op.test(n, target)))),
+        FieldValue::Range { low, high } => match (to_base(low, units), to_base(high, units)) {
+            (Some(low), Some(high)) => {
+                any(&mut |v| v.as_number().is_some_and(|n| n >= low && n <= high))
+            }
+            _ => false,
+        },
+    }
+}
+
+fn match_flag(any: &mut AnyValue, value: &FieldValue) -> bool {
+    let mut is_set = || any(&mut |v| v.as_flag() == Some(true));
+    match value {
+        FieldValue::Exists => is_set(),
+        FieldValue::Text { text, .. } => match parse_flag(text) {
+            Some(true) => is_set(),
+            Some(false) => !is_set(),
+            None => false,
+        },
+        _ => false,
+    }
+}
+
+/// Whether `pattern`, where `*` matches any run of characters, matches the whole of `text`.
+pub(crate) fn wildcard_match(pattern: &str, text: &str) -> bool {
+    let mut parts = pattern.split('*');
+    let first = parts.next().unwrap_or_default();
+    let Some(mut rest) = text.strip_prefix(first) else {
+        return false;
+    };
+    let parts: Vec<&str> = parts.collect();
+    let Some((last, middle)) = parts.split_last() else {
+        // No `*` at all: the whole text must equal the pattern
+        return rest.is_empty();
+    };
+    for part in middle {
+        match rest.find(part) {
+            Some(i) => rest = &rest[i + part.len()..],
+            None => return false,
+        }
+    }
+    rest.len() >= last.len() && rest.ends_with(last)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::schema::{FieldDef, Unit};
+    use crate::{format_field, parse};
+
+    fn text_field(name: &str) -> FieldDef {
+        FieldDef { name: name.into(), kind: FieldKind::Text, map: false, info: None }
+    }
+
+    /// The fields and free-text mode of Yaak's sidebar.
+    fn sidebar() -> Schema {
+        Schema::new(
+            [
+                "method",
+                "url",
+                "name",
+                "type",
+                "folder",
+                "grpc_service",
+                "grpc_method",
+            ]
+            .into_iter()
+            .map(text_field)
+            .collect(),
+            FreeText::Fuzzy,
+        )
+    }
+
+    fn proxy_like() -> Schema {
+        Schema::new(
+            vec![
+                text_field("url"),
+                FieldDef { name: "header".into(), kind: FieldKind::Text, map: true, info: None },
+                FieldDef {
+                    name: "status".into(),
+                    kind: FieldKind::Number { units: vec![] },
+                    map: false,
+                    info: None,
+                },
+                FieldDef {
+                    name: "duration".into(),
+                    kind: FieldKind::Number {
+                        units: vec![
+                            Unit { name: "ms".into(), factor: 1.0 },
+                            Unit { name: "s".into(), factor: 1000.0 },
+                        ],
+                    },
+                    map: false,
+                    info: None,
+                },
+                FieldDef { name: "error".into(), kind: FieldKind::Flag, map: false, info: None },
+            ],
+            FreeText::Substring,
+        )
+    }
+
+    /// An HTTP request in Yaak's sidebar, built the way `getItemFields` and `getItemText` do.
+    fn request(method: &str, name: &str, url: &str, folder: Option<&str>) -> Document {
+        let mut doc = Document::new(format!("{method} {name}"))
+            .with("method", method.to_uppercase())
+            .with("url", url)
+            .with("name", name)
+            .with("type", "http");
+        if let Some(folder) = folder {
+            doc = doc.with("folder", folder);
+        }
+        doc
+    }
+
+    fn grpc(name: &str, service: &str, method: &str) -> Document {
+        Document::new(name)
+            .with("grpc_method", method)
+            .with("grpc_service", service)
+            .with("url", "localhost:50051")
+            .with("name", name)
+            .with("type", "grpc")
+    }
+
+    fn matches(query: &str, schema: &Schema, doc: &Document) -> bool {
+        evaluate(parse(query).as_ref(), schema, doc)
+    }
+
+    fn sidebar_matches(query: &str, doc: &Document) -> bool {
+        matches(query, &sidebar(), doc)
+    }
+
+    #[test]
+    fn empty_query_matches_everything() {
+        assert!(sidebar_matches("", &Document::default()));
+        assert!(sidebar_matches("  ", &Document::default()));
+    }
+
+    #[test]
+    fn bare_words_are_fuzzy_in_the_sidebar() {
+        let doc = request("GET", "My Request", "https://yaak.app", None);
+        assert!(sidebar_matches("my", &doc));
+        assert!(sidebar_matches("REQ", &doc));
+        assert!(sidebar_matches("mr", &doc));
+        assert!(sidebar_matches("get", &doc));
+        assert!(!sidebar_matches("rq", &doc));
+        // Bare words match the free text only, not the URL
+        assert!(!sidebar_matches("yaak", &doc));
+    }
+
+    #[test]
+    fn bare_words_are_substrings_in_substring_mode() {
+        let doc = Document::new("GET /v1/charges 200");
+        let schema = proxy_like();
+        assert!(matches("rges", &schema, &doc));
+        assert!(matches("v1 charges", &schema, &doc));
+        assert!(!matches("gv", &schema, &doc));
+    }
+
+    #[test]
+    fn phrases_are_case_insensitive_substrings() {
+        let doc = request("GET", "My Request", "", None);
+        assert!(sidebar_matches("\"my req\"", &doc));
+        assert!(sidebar_matches("\"MY REQ\"", &doc));
+        assert!(!sidebar_matches("\"mr\"", &doc));
+        assert!(!sidebar_matches("\"\"", &doc));
+    }
+
+    #[test]
+    fn fields() {
+        let post = request("post", "Create user", "https://api.yaak.app/users", Some("Users"));
+        let get = request("get", "List users", "https://api.yaak.app/users", Some("Admin"));
+        let rpc = grpc("Say hello", "helloworld.Greeter", "SayHello");
+
+        assert!(sidebar_matches("@method:post", &post));
+        assert!(sidebar_matches("@METHOD:Po", &post));
+        assert!(!sidebar_matches("@method:post", &get));
+        assert!(sidebar_matches("@url:yaak.app/users", &post));
+        assert!(sidebar_matches("@folder:users", &post));
+        assert!(!sidebar_matches("@folder:users", &get));
+        assert!(sidebar_matches("@folder:users OR @folder:admin", &get));
+        assert!(sidebar_matches("@type:grpc", &rpc));
+        assert!(!sidebar_matches("-@type:grpc", &rpc));
+        assert!(sidebar_matches("-@type:grpc", &post));
+        assert!(sidebar_matches("@grpc_service:greeter @grpc_method:hello", &rpc));
+        // A field the document lacks matches nothing, and its negation everything
+        assert!(!sidebar_matches("@grpc_method:hello", &post));
+        assert!(sidebar_matches("-@grpc_method:hello", &post));
+        // `@` is required: without it, the term is free text
+        assert!(!sidebar_matches("method:post", &post));
+        assert!(sidebar_matches("localhost:50051", &Document::new("localhost:50051")));
+        // An unknown field matches nothing
+        assert!(!sidebar_matches("@nope:post", &post));
+        // A trailing field with no value yet matches nothing, as before
+        assert!(!sidebar_matches("@method:", &post));
+    }
+
+    #[test]
+    fn groups_and_negation() {
+        let a = Document::new("alpha");
+        let b = Document::new("beta");
+        let ac = Document::new("alpha charlie");
+        assert!(sidebar_matches("(alpha OR beta) -charlie", &a));
+        assert!(sidebar_matches("(alpha OR beta) -charlie", &b));
+        assert!(!sidebar_matches("(alpha OR beta) -charlie", &ac));
+        assert!(sidebar_matches("-(alpha OR beta)", &Document::new("delta")));
+        assert!(sidebar_matches("alpha charlie", &ac));
+        assert!(!sidebar_matches("alpha charlie", &a));
+    }
+
+    #[test]
+    fn wildcards() {
+        assert!(wildcard_match("*.github.com", "api.github.com"));
+        assert!(!wildcard_match("*.github.com", "github.com"));
+        assert!(wildcard_match("a*", "a"));
+        assert!(wildcard_match("*", ""));
+        assert!(wildcard_match("a*b*c", "aXXbYYc"));
+        assert!(!wildcard_match("a*b*c", "aXXbYYcd"));
+        assert!(wildcard_match("a**c", "ac"));
+        assert!(!wildcard_match("ab*ba", "aba"));
+
+        let doc = Document::default().with("url", "https://API.github.com/repos");
+        let schema = proxy_like();
+        assert!(matches("@url:https://api.github.com/*", &schema, &doc));
+        assert!(matches("@url:*GITHUB*", &schema, &doc));
+        assert!(!matches("@url:github*", &schema, &doc));
+        // Quoted, `*` is literal
+        assert!(!matches("@url:\"*github*\"", &schema, &doc));
+        assert!(matches("@url:\"*\"", &schema, &Document::default().with("url", "a*b")));
+    }
+
+    #[test]
+    fn presence() {
+        let schema = proxy_like();
+        let with = Document::default().with("header.x-cache", "HIT").with("status", 200.0);
+        let without = Document::default().with("url", "x").with("status", 200.0);
+        assert!(matches("@header.x-cache:*", &schema, &with));
+        assert!(!matches("@header.x-cache:*", &schema, &without));
+        assert!(matches("-@header.x-cache:*", &schema, &without));
+        assert!(matches("@status:*", &schema, &with));
+        assert!(!matches("@url:*", &schema, &Document::default().with("url", "")));
+    }
+
+    #[test]
+    fn map_fields() {
+        let schema = proxy_like();
+        let doc = Document::default()
+            .with("header.content-type", "application/json")
+            .with("header.x-cache", "HIT");
+        assert!(matches("@header.content-type:json", &schema, &doc));
+        assert!(matches("@header.Content-Type:JSON", &schema, &doc));
+        assert!(!matches("@header.content-type:xml", &schema, &doc));
+        assert!(!matches("@header.accept:json", &schema, &doc));
+        // The map with no key looks at every key
+        assert!(matches("@header:hit", &schema, &doc));
+        assert!(matches("@header:*", &schema, &doc));
+    }
+
+    #[test]
+    fn numbers() {
+        let schema = proxy_like();
+        let doc = Document::default().with("status", 404.0).with("duration", 1500.0);
+        assert!(matches("@status:404", &schema, &doc));
+        assert!(!matches("@status:40", &schema, &doc));
+        assert!(matches("@status:40*", &schema, &doc));
+        assert!(matches("@status:>=400", &schema, &doc));
+        assert!(!matches("@status:<400", &schema, &doc));
+        assert!(matches("@status:[400 TO 499]", &schema, &doc));
+        assert!(matches("@status:[404 TO 404]", &schema, &doc));
+        assert!(!matches("@status:[200 TO 299]", &schema, &doc));
+        assert!(matches("@duration:>1s", &schema, &doc));
+        assert!(matches("@duration:>1000ms", &schema, &doc));
+        assert!(matches("@duration:1.5s", &schema, &doc));
+        assert!(matches("@duration:[1s TO 2s]", &schema, &doc));
+        assert!(!matches("@duration:>2s", &schema, &doc));
+        // A unit the field does not take matches nothing
+        assert!(!matches("@duration:>1m", &schema, &doc));
+        assert!(!matches("@status:>1ms", &schema, &doc));
+    }
+
+    #[test]
+    fn comparisons_on_text_fields() {
+        let doc = Document::default().with("url", "404");
+        assert!(matches("@url:>400", &proxy_like(), &doc));
+        assert!(!matches("@url:>400", &proxy_like(), &Document::default().with("url", "x")));
+    }
+
+    #[test]
+    fn flags() {
+        let schema = proxy_like();
+        let failed = Document::default().with("error", true);
+        let ok = Document::default().with("error", false);
+        assert!(matches("@error:*", &schema, &failed));
+        assert!(!matches("@error:*", &schema, &ok));
+        assert!(matches("@error:true", &schema, &failed));
+        assert!(matches("@error:false", &schema, &ok));
+        assert!(matches("@error:false", &schema, &Document::default()));
+        assert!(!matches("@error:maybe", &schema, &failed));
+    }
+
+    #[test]
+    fn multiple_values() {
+        let doc = Document::default().with("url", "a").with("url", "b");
+        assert!(matches("@url:b", &proxy_like(), &doc));
+    }
+
+    /// A table gives the same answers as matching its rows one document at a time.
+    #[test]
+    fn tables_match_like_documents() {
+        let columns = [
+            "method",
+            "url",
+            "name",
+            "type",
+            "folder",
+            "grpc_service",
+            "grpc_method",
+        ];
+        let docs = [
+            request("post", "Create user", "https://api.yaak.app/users", Some("Users")),
+            request("get", "List users", "https://api.yaak.app/users", Some("Admin")),
+            request("GET", "My Request", "https://yaak.app", None),
+            grpc("Say hello", "helloworld.Greeter", "SayHello"),
+            Document::new("Folder"),
+        ];
+        let mut table = Table::new(columns);
+        for doc in &docs {
+            let cells = columns.iter().map(|c| match doc.fields.get(*c) {
+                Some(DocValues::One(DocValue::Text(v))) => Some(Cow::Borrowed(v.as_str())),
+                _ => None,
+            });
+            table.push(doc.text.as_str(), cells);
+        }
+        assert_eq!(table.len(), docs.len());
+
+        for query in [
+            "",
+            "user",
+            "mr",
+            "\"my req\"",
+            "@method:post",
+            "-@type:grpc",
+            "@folder:users OR @folder:admin",
+            "(users OR hello) -create",
+            "@grpc_method:*",
+            "-@folder:*",
+            "@url:*yaak.app*",
+            "@nope:x",
+            "-@nope:x",
+            "method:post",
+        ] {
+            let matcher = Matcher::new(parse(query).as_ref(), &sidebar());
+            let expected: Vec<u32> = (0..docs.len() as u32)
+                .filter(|&i| sidebar_matches(query, &docs[i as usize]))
+                .collect();
+            assert_eq!(matcher.matching_rows(&table), expected, "{query:?}");
+        }
+        assert!(Matcher::new(None, &sidebar()).is_empty());
+    }
+
+    #[test]
+    fn table_cells_line_up_with_columns() {
+        let mut table = Table::new(["Method", "url"]);
+        table.push("a", [Some("GET".into())]);
+        table.push("b", [None, Some("x".into()), Some("ignored".into())]);
+        assert_eq!(table.len(), 2);
+        let rows = |q: &str| Matcher::new(parse(q).as_ref(), &sidebar()).matching_rows(&table);
+        assert_eq!(rows("@method:get"), vec![0]);
+        assert_eq!(rows("@url:x"), vec![1]);
+        assert_eq!(rows("-@url:*"), vec![0]);
+        assert_eq!(rows("b"), vec![1]);
+    }
+
+    #[test]
+    fn tables_from_flat_cells() {
+        let cells = vec![
+            Some("GET List".into()),
+            Some("GET".into()),
+            None,
+            Some("POST".into()),
+        ];
+        let table = Table::from_cells(["method"], cells.clone()).unwrap();
+        assert_eq!(table.len(), 2);
+        let rows = |q: &str| Matcher::new(parse(q).as_ref(), &sidebar()).matching_rows(&table);
+        assert_eq!(rows("@method:post"), vec![1]);
+        assert_eq!(rows("list"), vec![0]);
+        // Cells that don't fill whole rows
+        assert!(Table::from_cells(["method", "url"], cells).is_none());
+        assert!(Table::from_cells(["method"], Vec::new()).unwrap().is_empty());
+    }
+
+    /// Packs cells the way the JavaScript side does: joined, with UTF-16 lengths.
+    fn pack(cells: &[Option<&str>]) -> (String, Vec<u32>) {
+        let joined = cells.iter().flatten().copied().collect();
+        let lengths = cells
+            .iter()
+            .map(|c| c.map_or(Table::NO_VALUE, |c| c.encode_utf16().count() as u32))
+            .collect();
+        (joined, lengths)
+    }
+
+    #[test]
+    fn tables_from_packed_cells() {
+        let rows = |table: &Table, q: &str| {
+            Matcher::new(parse(q).as_ref(), &sidebar()).matching_rows(table)
+        };
+
+        let (joined, lengths) = pack(&[
+            Some("GET List"),
+            Some("GET"),
+            Some(""),
+            None,
+            Some("POST"),
+            Some("x"),
+        ]);
+        let table = Table::from_packed(["method", "url"], &joined, &lengths).unwrap();
+        assert_eq!(table.len(), 2);
+        assert_eq!(rows(&table, "@method:post"), vec![1]);
+        assert_eq!(rows(&table, "@url:x"), vec![1]);
+        assert_eq!(rows(&table, "list"), vec![0]);
+        assert_eq!(rows(&table, "-@method:*"), Vec::<u32>::new());
+
+        // Non-ASCII text, including a character outside the BMP (two UTF-16 units)
+        let (joined, lengths) = pack(&[
+            Some("Über 🦀 crab"),
+            Some("GET"),
+            Some("Café"),
+            Some("x"),
+            None,
+            Some("y"),
+        ]);
+        let table = Table::from_packed(["method", "name"], &joined, &lengths).unwrap();
+        assert_eq!(rows(&table, "🦀"), vec![0]);
+        assert_eq!(rows(&table, "@name:café"), vec![0]);
+        assert_eq!(rows(&table, "@name:y"), vec![1]);
+
+        // Lengths that don't add up, or don't fill whole rows
+        assert!(Table::from_packed(["method"], "abc", &[1, 1]).is_none());
+        assert!(Table::from_packed(["method"], "abc", &[2, 2]).is_none());
+        assert!(Table::from_packed(["method"], "abc", &[3]).is_none());
+        assert!(Table::from_packed(["method"], "é", &[2, 0]).is_none());
+        assert!(Table::from_packed(["method"], "", &[]).unwrap().is_empty());
+    }
+
+    #[test]
+    fn non_ascii_text_matches_case_insensitively() {
+        let doc = Document::new("ÜBER Café").with("name", "ÜBER Café");
+        assert!(sidebar_matches("über", &doc));
+        assert!(sidebar_matches("@name:café", &doc));
+        assert!(sidebar_matches("\"über c\"", &doc));
+        assert!(!sidebar_matches("@name:cafe", &doc));
+    }
+
+    #[test]
+    fn document_from_json() {
+        let doc: Document = serde_json::from_str(
+            r#"{ "text": "x", "fields": { "url": "a", "tags": ["b", "c"], "status": 200, "error": true } }"#,
+        )
+        .unwrap();
+        assert_eq!(doc.fields["url"], DocValues::One(DocValue::Text("a".into())));
+        assert_eq!(doc.fields["status"], DocValues::One(DocValue::Number(200.0)));
+        assert_eq!(doc.fields["error"], DocValues::One(DocValue::Flag(true)));
+        assert!(matches("@status:200 @error:true", &proxy_like(), &doc));
+    }
+
+    // The cases from Yaak's former `format.test.ts`: a formatted field filter matches the value
+    // it was made from. Values that contain query syntax are now quoted so they stay literal.
+
+    fn matches_formatted_url(value: &str) -> bool {
+        matches(
+            &format_field("url", value, false),
+            &sidebar(),
+            &Document::default().with("url", value),
+        )
+    }
+
+    #[test]
+    fn format_keeps_url_like_values_bare() {
+        assert_eq!(format_field("url", "yaak.app/foo-bar", false), "@url:yaak.app/foo-bar");
+        assert!(matches_formatted_url("yaak.app/foo-bar"));
+    }
+
+    #[test]
+    fn format_quotes_values_with_query_syntax() {
+        // `*` is a wildcard and `)` closes a group, so this one is quoted now
+        assert_eq!(format_field("url", "yaa$&#*@tsrna(*)", false), "@url:\"yaa$&#*@tsrna(*)\"");
+        assert!(matches_formatted_url("yaa$&#*@tsrna(*)"));
+        assert_eq!(format_field("url", "yaa$&#@tsrna(", false), "@url:yaa$&#@tsrna(");
+        assert!(matches_formatted_url("yaa$&#@tsrna("));
+    }
+
+    #[test]
+    fn format_keeps_values_that_start_with_an_operator_token_bare() {
+        assert_eq!(format_field("url", "-foo", false), "@url:-foo");
+        assert!(matches_formatted_url("-foo"));
+    }
+
+    #[test]
+    fn format_keeps_boolean_operator_words_bare() {
+        assert_eq!(format_field("url", "AND", false), "@url:AND");
+        assert_eq!(format_field("url", "or", false), "@url:or");
+        assert_eq!(format_field("url", "Not", false), "@url:Not");
+        assert!(matches_formatted_url("AND"));
+    }
+
+    #[test]
+    fn format_escapes_quoted_values() {
+        assert_eq!(format_field("url", "say \"hi\"", false), "@url:\"say \\\"hi\\\"\"");
+        assert!(matches_formatted_url("say \"hi\""));
+    }
+
+    #[test]
+    fn format_quotes_values_that_start_with_a_quote() {
+        assert_eq!(format_field("url", "\"hi\"", false), "@url:\"\\\"hi\\\"\"");
+        assert!(matches_formatted_url("\"hi\""));
+    }
+
+    #[test]
+    fn formatted_values_round_trip() {
+        for value in [
+            "a b",
+            "*",
+            ">5",
+            "[1 TO 2]",
+            "x)",
+            "back\\slash",
+            "\\\"",
+            "tab\there",
+            "é ü",
+            "a*b",
+            "<=3ms",
+            "\\",
+        ] {
+            assert!(matches_formatted_url(value), "{value:?} did not round-trip");
+        }
+    }
+}
diff --git a/crates/yaak-query/src/format.rs b/crates/yaak-query/src/format.rs
new file mode 100644
index 000000000..a6cc2d8ba
--- /dev/null
+++ b/crates/yaak-query/src/format.rs
@@ -0,0 +1,89 @@
+//! Writes field terms back out as query text.
+
+use crate::ast::FieldValue;
+use crate::lexer::{classify_bare, is_ident_char};
+
+/// A term matching `value` in `field`: `@field:value`, or `-@field:value` when `negated`. The
+/// value stays bare when it would read back as the same literal text, and is quoted (with `"`
+/// and `\` escaped) when it holds whitespace, a quote, a `)`, a `*`, or the shape of a
+/// comparison or range.
+pub fn format_field(field: &str, value: &str, negated: bool) -> String {
+    let sign = if negated { "-" } else { "" };
+    format!("{sign}@{field}:{}", format_value(value))
+}
+
+/// A field value as it should be written after the colon.
+pub fn format_value(value: &str) -> String {
+    if is_bare_value(value) { value.to_string() } else { quote(value) }
+}
+
+/// `value` in double quotes, with `"` and `\` escaped.
+pub fn quote(value: &str) -> String {
+    let mut out = String::with_capacity(value.len() + 2);
+    out.push('"');
+    for c in value.chars() {
+        if matches!(c, '"' | '\\') {
+            out.push('\\');
+        }
+        out.push(c);
+    }
+    out.push('"');
+    out
+}
+
+fn is_bare_value(value: &str) -> bool {
+    let Some(first) = value.chars().next() else {
+        return false;
+    };
+    if first == '"' || first == '[' {
+        return false;
+    }
+    if value.chars().any(|c| c.is_whitespace() || c == ')') {
+        return false;
+    }
+    matches!(classify_bare(value), FieldValue::Text { .. })
+}
+
+/// Whether `name` can be written as a field path: identifiers joined by dots.
+pub fn is_valid_path(name: &str) -> bool {
+    !name.is_empty()
+        && name.split('.').all(|part| !part.is_empty() && part.chars().all(is_ident_char))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn bare_or_quoted() {
+        assert_eq!(format_value("post"), "post");
+        assert_eq!(format_value("a b"), "\"a b\"");
+        assert_eq!(format_value(""), "\"\"");
+        assert_eq!(format_value("*"), "\"*\"");
+        assert_eq!(format_value("a*"), "\"a*\"");
+        assert_eq!(format_value(">5"), "\">5\"");
+        assert_eq!(format_value(">x"), ">x");
+        assert_eq!(format_value("[1 TO 2]"), "\"[1 TO 2]\"");
+        assert_eq!(format_value("[x"), "\"[x\"");
+        assert_eq!(format_value("a)"), "\"a)\"");
+        assert_eq!(format_value("a(b"), "a(b");
+        assert_eq!(format_value("a\"b"), "a\"b");
+        assert_eq!(format_value("a\\b"), "a\\b");
+        assert_eq!(format_value("x\\ y"), "\"x\\\\ y\"");
+    }
+
+    #[test]
+    fn negated() {
+        assert_eq!(format_field("type", "grpc", true), "-@type:grpc");
+        assert_eq!(format_field("name", "My Request", true), "-@name:\"My Request\"");
+    }
+
+    #[test]
+    fn paths() {
+        assert!(is_valid_path("method"));
+        assert!(is_valid_path("req.header.x-cache"));
+        assert!(!is_valid_path(""));
+        assert!(!is_valid_path("a..b"));
+        assert!(!is_valid_path("a b"));
+    }
+}
diff --git a/crates/yaak-query/src/fuzzy.rs b/crates/yaak-query/src/fuzzy.rs
new file mode 100644
index 000000000..153606cc8
--- /dev/null
+++ b/crates/yaak-query/src/fuzzy.rs
@@ -0,0 +1,206 @@
+//! Word-prefix matching, ported from fuzzbunny's `fuzzyMatchSanitized` so Yaak's sidebar keeps
+//! the behaviour it had when it called fuzzbunny directly.
+//!
+//! A search matches when it is a substring of the target. Failing that, it matches when it can
+//! be split into runs that each start a word or sit at a punctuation mark, in order: `mr` finds
+//! "My Request", `usam` finds "United States of America", but `rq` does not find "request".
+//! Spaces in either string never break a run.
+//!
+//! Only whether it matches is kept; fuzzbunny's scores and highlight ranges are not used.
+
+/// Whether `search` matches `target`, both already lowercased and trimmed.
+pub fn fuzzy_match(target: &str, search: &str) -> bool {
+    fuzzy_match_in(target, search, &mut FuzzyScratch::default())
+}
+
+/// Buffers [`fuzzy_match_in`] reuses from one call to the next.
+#[derive(Default)]
+pub(crate) struct FuzzyScratch {
+    target: Vec<char>,
+    search: Vec<char>,
+    skips: Vec<usize>,
+}
+
+/// [`fuzzy_match`], reusing `scratch` instead of allocating per call. ASCII text, the usual case,
+/// is matched byte by byte without copying.
+pub(crate) fn fuzzy_match_in(target: &str, search: &str, scratch: &mut FuzzyScratch) -> bool {
+    if target.is_empty() {
+        return false;
+    }
+    if search.is_empty() {
+        return true;
+    }
+
+    // A search starting with a quote wants a substring only; the closing quote is optional
+    let quoted = search.starts_with('"');
+    let search = if quoted {
+        let inner = &search[1..];
+        inner.strip_suffix('"').unwrap_or(inner)
+    } else {
+        search
+    };
+
+    if target.contains(search) {
+        return true;
+    }
+
+    // One character that is not a substring can't match fuzzily either
+    let mut chars = search.chars();
+    if quoted || (chars.next().is_some() && chars.next().is_none()) {
+        return false;
+    }
+
+    let skips = &mut scratch.skips;
+    if target.is_ascii() && search.is_ascii() {
+        return word_prefix_match(target.as_bytes(), search.as_bytes(), skips);
+    }
+    scratch.target.clear();
+    scratch.target.extend(target.chars());
+    scratch.search.clear();
+    scratch.search.extend(search.chars());
+    word_prefix_match(&scratch.target, &scratch.search, skips)
+}
+
+/// A unit of text the matcher walks: a byte of ASCII text, or a char.
+trait Unit: Copy + PartialEq {
+    /// fuzzbunny counts lowercase letters, digits, and every char code from 127 up.
+    fn is_alpha_num(self) -> bool;
+    /// ASCII punctuation: `!` to `/`, `:` to `@`, `[` to `` ` ``, `{` to `~`.
+    fn is_punctuation(self) -> bool;
+    fn is_space(self) -> bool;
+}
+
+impl Unit for char {
+    fn is_alpha_num(self) -> bool {
+        self.is_ascii_lowercase() || self.is_ascii_digit() || self as u32 >= 127
+    }
+
+    fn is_punctuation(self) -> bool {
+        matches!(self, '!'..='/' | ':'..='@' | '['..='`' | '{'..='~')
+    }
+
+    fn is_space(self) -> bool {
+        self == ' '
+    }
+}
+
+impl Unit for u8 {
+    fn is_alpha_num(self) -> bool {
+        (self as char).is_alpha_num()
+    }
+
+    fn is_punctuation(self) -> bool {
+        (self as char).is_punctuation()
+    }
+
+    fn is_space(self) -> bool {
+        self == b' '
+    }
+}
+
+/// The word-prefix fallback, once `search` is known not to be a plain substring.
+fn word_prefix_match<T: Unit>(target: &[T], search: &[T], skips: &mut Vec<usize>) -> bool {
+    target_skips(target, skips);
+    (0..skips.len() - 1)
+        .any(|skip| target[skips[skip]] == search[0] && prefix_match(skip, search, target, skips))
+}
+
+/// The indices where a word starts or a punctuation mark sits, then the target's length, so
+/// every run lies between `skips[i]` and `skips[i + 1]`.
+fn target_skips<T: Unit>(target: &[T], skips: &mut Vec<usize>) {
+    skips.clear();
+    let mut was_alpha_num = false;
+    for (i, &c) in target.iter().enumerate() {
+        let alpha_num = c.is_alpha_num();
+        if (alpha_num && !was_alpha_num) || c.is_punctuation() {
+            skips.push(i);
+        }
+        was_alpha_num = alpha_num;
+    }
+    skips.push(target.len());
+}
+
+fn prefix_match<T: Unit>(mut skip: usize, search: &[T], target: &[T], skips: &[usize]) -> bool {
+    let mut search_idx = 0;
+    while skip < skips.len() - 1 {
+        let mut target_idx = skips[skip];
+        let end = skips[skip + 1];
+        while target_idx < end && search_idx < search.len() {
+            let t = target[target_idx];
+            let s = search[search_idx];
+            if t == s {
+                target_idx += 1;
+                search_idx += 1;
+            } else if t.is_space() {
+                target_idx += 1;
+            } else if s.is_space() {
+                search_idx += 1;
+            } else {
+                break;
+            }
+        }
+        if search_idx == search.len() {
+            return true;
+        }
+        skip += 1;
+    }
+    false
+}
+
+#[cfg(test)]
+mod tests {
+    use super::fuzzy_match;
+
+    fn m(target: &str, search: &str) -> bool {
+        fuzzy_match(&target.to_lowercase(), &search.trim().to_lowercase())
+    }
+
+    #[test]
+    fn substring_first() {
+        assert!(m("My Request", "req"));
+        assert!(m("My Request", "quest"));
+        assert!(m("/v1/charges", "rges"));
+        assert!(m("abc", "abc"));
+    }
+
+    #[test]
+    fn word_prefixes() {
+        assert!(m("My Request", "mr"));
+        assert!(m("My Request", "myreq"));
+        assert!(m("United States of America", "usam"));
+        assert!(m("api.github.com", "apigi"));
+        // Each run must continue from the start of a word: "gh" is not a prefix of "github"
+        assert!(!m("api.github.com", "apigh"));
+        assert!(m("POST Create User", "postcu"));
+        assert!(m("get_user_by_id", "gubi"));
+    }
+
+    #[test]
+    fn not_any_subsequence() {
+        assert!(!m("My Request", "rq"));
+        assert!(!m("request", "rqst"));
+        assert!(!m("My Request", "mx"));
+        assert!(!m("abc", "x"));
+    }
+
+    #[test]
+    fn spaces_do_not_break_a_run() {
+        assert!(m("My Request", "my r"));
+        assert!(m("foo bar", "f b"));
+    }
+
+    #[test]
+    fn empty_and_quoted() {
+        assert!(!m("", "a"));
+        assert!(m("abc", ""));
+        assert!(m("las vegas", "\"la"));
+        assert!(!m("los angeles", "\"la\""));
+        assert!(m("los angeles", "la"));
+    }
+
+    #[test]
+    fn unicode_is_alphanumeric() {
+        assert!(m("Über Café", "üc"));
+        assert!(m("日本 語", "日語"));
+    }
+}
diff --git a/crates/yaak-query/src/lexer.rs b/crates/yaak-query/src/lexer.rs
new file mode 100644
index 000000000..b28ae24ee
--- /dev/null
+++ b/crates/yaak-query/src/lexer.rs
@@ -0,0 +1,320 @@
+//! Splits a query into tokens. Never fails: anything it does not recognise is text.
+
+use crate::ast::{CompareOp, FieldValue, NumberLiteral, Span};
+
+#[derive(Clone, Debug, PartialEq)]
+pub(crate) enum TokKind {
+    LParen,
+    RParen,
+    /// `-` at the start of a term. It negates only what follows it directly.
+    Minus,
+    Or,
+    And,
+    Word(String),
+    Phrase {
+        text: String,
+    },
+    Field {
+        path: Vec<String>,
+        path_span: Span,
+        value: FieldValue,
+        value_span: Span,
+    },
+}
+
+#[derive(Clone, Debug, PartialEq)]
+pub(crate) struct Tok {
+    pub kind: TokKind,
+    pub span: Span,
+}
+
+pub(crate) fn lex(input: &str) -> Vec<Tok> {
+    let mut toks = Vec::new();
+    let mut i = 0;
+    while let Some(c) = input[i..].chars().next() {
+        if c.is_whitespace() {
+            i += c.len_utf8();
+            continue;
+        }
+        let (kind, end) = match c {
+            '(' => (TokKind::LParen, i + 1),
+            ')' => (TokKind::RParen, i + 1),
+            '-' => (TokKind::Minus, i + 1),
+            '"' => {
+                let (text, end) = lex_quoted(input, i);
+                (TokKind::Phrase { text }, end)
+            }
+            '@' => match lex_field(input, i) {
+                Some(field) => field,
+                None => lex_word(input, i),
+            },
+            _ => lex_word(input, i),
+        };
+        toks.push(Tok { kind, span: Span::new(i, end) });
+        i = end;
+    }
+    toks
+}
+
+/// Whether `c` ends a bare free-text word.
+fn ends_word(c: char) -> bool {
+    c.is_whitespace() || matches!(c, '(' | ')' | '"')
+}
+
+/// Whether `c` ends a bare field value. Parentheses inside a value are kept, except a closing
+/// one, which closes the group the field sits in.
+fn ends_value(c: char) -> bool {
+    c.is_whitespace() || c == ')'
+}
+
+fn run_until(input: &str, start: usize, stop: impl Fn(char) -> bool) -> usize {
+    input[start..].find(stop).map_or(input.len(), |n| start + n)
+}
+
+fn lex_word(input: &str, start: usize) -> (TokKind, usize) {
+    // The first character is never a word end (the caller checked), so a word is never empty
+    let first = input[start..].chars().next().map_or(0, char::len_utf8);
+    let end = run_until(input, start + first, ends_word);
+    let text = &input[start..end];
+    let kind = match text {
+        "OR" => TokKind::Or,
+        "AND" => TokKind::And,
+        _ => TokKind::Word(text.to_string()),
+    };
+    (kind, end)
+}
+
+/// A double-quoted string starting at `start`. `\"` and `\\` are escapes; any other backslash
+/// is kept. Without a closing quote it runs to the end of the input.
+pub(crate) fn lex_quoted(input: &str, start: usize) -> (String, usize) {
+    let mut text = String::new();
+    let mut chars = input[start + 1..].char_indices();
+    while let Some((n, c)) = chars.next() {
+        match c {
+            '"' => return (text, start + 1 + n + 1),
+            '\\' => match input[start + 1 + n + 1..].chars().next() {
+                Some(next @ ('"' | '\\')) => {
+                    text.push(next);
+                    chars.next();
+                }
+                _ => text.push('\\'),
+            },
+            c => text.push(c),
+        }
+    }
+    (text, input.len())
+}
+
+pub(crate) fn is_ident_char(c: char) -> bool {
+    c.is_ascii_alphanumeric() || c == '_' || c == '-'
+}
+
+/// `@path:value`, or `None` when the text at `start` is not one (no colon after the path).
+fn lex_field(input: &str, start: usize) -> Option<(TokKind, usize)> {
+    let mut path = Vec::new();
+    let mut i = start + 1;
+    loop {
+        let end = run_until(input, i, |c| !is_ident_char(c));
+        if end == i {
+            return None;
+        }
+        path.push(input[i..end].to_string());
+        i = end;
+        match input[i..].chars().next() {
+            Some('.') => i += 1,
+            Some(':') => break,
+            _ => return None,
+        }
+    }
+    let path_span = Span::new(start, i + 1);
+    let (value, value_end) = lex_value(input, i + 1);
+    let kind = TokKind::Field { path, path_span, value, value_span: Span::new(i + 1, value_end) };
+    Some((kind, value_end))
+}
+
+fn lex_value(input: &str, start: usize) -> (FieldValue, usize) {
+    match input[start..].chars().next() {
+        None => return (FieldValue::Text { text: String::new(), quoted: false }, start),
+        Some(c) if ends_value(c) => {
+            return (FieldValue::Text { text: String::new(), quoted: false }, start);
+        }
+        Some('"') => {
+            let (text, end) = lex_quoted(input, start);
+            return (FieldValue::Text { text, quoted: true }, end);
+        }
+        Some('[') => {
+            if let Some((low, high, end)) = lex_range(input, start) {
+                return (FieldValue::Range { low, high }, end);
+            }
+        }
+        Some(_) => {}
+    }
+    let end = run_until(input, start, ends_value);
+    (classify_bare(&input[start..end]), end)
+}
+
+/// What a bare (unquoted) field value means: presence, a comparison, a wildcard pattern, or
+/// plain text.
+pub(crate) fn classify_bare(raw: &str) -> FieldValue {
+    if raw == "*" {
+        return FieldValue::Exists;
+    }
+    if let Some((op, number)) = parse_compare(raw) {
+        return FieldValue::Compare { op, number };
+    }
+    if raw.contains('*') {
+        return FieldValue::Wildcard { pattern: raw.to_string() };
+    }
+    FieldValue::Text { text: raw.to_string(), quoted: false }
+}
+
+fn parse_compare(raw: &str) -> Option<(CompareOp, NumberLiteral)> {
+    let (op, rest) = if let Some(rest) = raw.strip_prefix(">=") {
+        (CompareOp::Gte, rest)
+    } else if let Some(rest) = raw.strip_prefix("<=") {
+        (CompareOp::Lte, rest)
+    } else if let Some(rest) = raw.strip_prefix('>') {
+        (CompareOp::Gt, rest)
+    } else if let Some(rest) = raw.strip_prefix('<') {
+        (CompareOp::Lt, rest)
+    } else {
+        return None;
+    };
+    Some((op, parse_number(rest)?))
+}
+
+/// A number with an optional unit: `-?digits(.digits)?letters?`, the whole of `raw`.
+pub(crate) fn parse_number(raw: &str) -> Option<NumberLiteral> {
+    let body = raw.strip_prefix('-').unwrap_or(raw);
+    let digits_end = body.find(|c: char| !c.is_ascii_digit()).unwrap_or(body.len());
+    if digits_end == 0 {
+        return None;
+    }
+    let mut end = digits_end;
+    if let Some(fraction) = body[end..].strip_prefix('.') {
+        let fraction_len = fraction.find(|c: char| !c.is_ascii_digit()).unwrap_or(fraction.len());
+        if fraction_len == 0 {
+            return None;
+        }
+        end += 1 + fraction_len;
+    }
+    let unit = &body[end..];
+    if !unit.chars().all(|c| c.is_ascii_alphabetic()) {
+        return None;
+    }
+    let number_len = raw.len() - body.len() + end;
+    let value: f64 = raw[..number_len].parse().ok()?;
+    Some(NumberLiteral {
+        value,
+        unit: (!unit.is_empty()).then(|| unit.to_string()),
+        raw: raw.to_string(),
+    })
+}
+
+/// `[low TO high]` starting at `start`, or `None` when the text does not have exactly that
+/// shape (it is then a literal value).
+fn lex_range(input: &str, start: usize) -> Option<(NumberLiteral, NumberLiteral, usize)> {
+    let skip_ws = |i: usize| run_until(input, i, |c| !c.is_whitespace());
+    let mut i = skip_ws(start + 1);
+    let low_end = run_until(input, i, |c| c.is_whitespace() || c == ']');
+    let low = parse_number(&input[i..low_end])?;
+    i = skip_ws(low_end);
+    if i == low_end {
+        return None;
+    }
+    input[i..].strip_prefix("TO")?;
+    i += 2;
+    let after_to = skip_ws(i);
+    if after_to == i {
+        return None;
+    }
+    i = after_to;
+    let high_end = run_until(input, i, |c| c.is_whitespace() || c == ']');
+    let high = parse_number(&input[i..high_end])?;
+    i = skip_ws(high_end);
+    input[i..].strip_prefix(']')?;
+    let end = i + 1;
+    match input[end..].chars().next() {
+        None => Some((low, high, end)),
+        Some(c) if ends_value(c) => Some((low, high, end)),
+        Some(_) => None,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn kinds(input: &str) -> Vec<TokKind> {
+        lex(input).into_iter().map(|t| t.kind).collect()
+    }
+
+    #[test]
+    fn numbers() {
+        let n = parse_number("500ms").unwrap();
+        assert_eq!(n.value, 500.0);
+        assert_eq!(n.unit.as_deref(), Some("ms"));
+        assert_eq!(parse_number("-2.5").unwrap().value, -2.5);
+        assert_eq!(parse_number("400").unwrap().unit, None);
+        assert!(parse_number("").is_none());
+        assert!(parse_number("ms").is_none());
+        assert!(parse_number("1.").is_none());
+        assert!(parse_number("1.5.2").is_none());
+        assert!(parse_number("12a3").is_none());
+        assert!(parse_number("-").is_none());
+    }
+
+    #[test]
+    fn words_and_operators() {
+        assert_eq!(
+            kinds("a OR b AND or"),
+            vec![
+                TokKind::Word("a".into()),
+                TokKind::Or,
+                TokKind::Word("b".into()),
+                TokKind::And,
+                TokKind::Word("or".into()),
+            ]
+        );
+    }
+
+    #[test]
+    fn words_stop_at_parens_and_quotes() {
+        assert_eq!(
+            kinds("(a)b\"c d\""),
+            vec![
+                TokKind::LParen,
+                TokKind::Word("a".into()),
+                TokKind::RParen,
+                TokKind::Word("b".into()),
+                TokKind::Phrase { text: "c d".into() },
+            ]
+        );
+    }
+
+    #[test]
+    fn quoted_escapes() {
+        assert_eq!(lex_quoted(r#""a \"b\" \\ \n""#, 0), (r#"a "b" \ \n"#.to_string(), 15));
+        assert_eq!(lex_quoted(r#""open"#, 0), ("open".to_string(), 5));
+        assert_eq!(lex_quoted(r#""trailing\"#, 0), ("trailing\\".to_string(), 10));
+    }
+
+    #[test]
+    fn spans_are_bytes() {
+        let toks = lex("é @url:ü");
+        assert_eq!(toks[0].span, Span::new(0, 2));
+        assert_eq!(toks[1].span, Span::new(3, 10));
+    }
+
+    #[test]
+    fn range_needs_its_whole_shape() {
+        assert!(lex_range("[1 TO 2]", 0).is_some());
+        assert!(lex_range("[ 1  TO  2 ]", 0).is_some());
+        assert!(lex_range("[1 to 2]", 0).is_none());
+        assert!(lex_range("[1TO 2]", 0).is_none());
+        assert!(lex_range("[1 TO 2", 0).is_none());
+        assert!(lex_range("[1 TO 2]x", 0).is_none());
+        assert!(lex_range("[a TO b]", 0).is_none());
+        assert!(lex_range("[ base_url ]", 0).is_none());
+    }
+}
diff --git a/crates/yaak-query/src/lib.rs b/crates/yaak-query/src/lib.rs
new file mode 100644
index 000000000..14cc87e66
--- /dev/null
+++ b/crates/yaak-query/src/lib.rs
@@ -0,0 +1,40 @@
+//! The search query language of Yaak and Yaak Proxy.
+//!
+//! The syntax follows Datadog's log search. Terms side by side must all match; `OR` (upper
+//! case) between terms or groups means either; parentheses group; `-` directly before a term or
+//! group excludes it. A bare word matches the free text, a double-quoted phrase matches it with
+//! its spaces kept, and `@path:value` matches a field. A value may be `*` (the field is present),
+//! a `*` wildcard pattern, a comparison like `>=400` or `>500ms`, or a range like
+//! `[400 TO 499]`. Everything matches case-insensitively, and nothing in the input is an error.
+//!
+//! [`parse`] gives the tree, which a consumer may walk to compile it elsewhere (to SQL, say).
+//! [`evaluate`] runs it in memory against a [`Document`], with a [`Schema`] saying what the
+//! fields are. [`tokens`], [`completions`] and [`format_field`] serve the search box.
+
+// The wasm-bindgen exports expand to unsafe glue; nothing else in the crate may use unsafe
+#![cfg_attr(
+    not(all(feature = "wasm", target_arch = "wasm32")),
+    forbid(unsafe_code)
+)]
+#![cfg_attr(all(feature = "wasm", target_arch = "wasm32"), deny(unsafe_code))]
+
+pub mod ast;
+mod complete;
+mod eval;
+mod format;
+pub mod fuzzy;
+mod lexer;
+mod parser;
+mod schema;
+mod tokens;
+#[cfg(all(feature = "wasm", target_arch = "wasm32"))]
+#[allow(unsafe_code)]
+mod wasm;
+
+pub use ast::{CompareOp, FieldValue, Node, NumberLiteral, Span};
+pub use complete::{Completion, CompletionKind, Completions, completions};
+pub use eval::{DocValue, DocValues, Document, Matcher, Table, evaluate};
+pub use format::{format_field, format_value, is_valid_path, quote};
+pub use parser::parse;
+pub use schema::{Diagnostic, FieldDef, FieldKind, FreeText, ResolvedField, Schema, Unit};
+pub use tokens::{Token, TokenKind, tokens};
diff --git a/crates/yaak-query/src/parser.rs b/crates/yaak-query/src/parser.rs
new file mode 100644
index 000000000..3ef505af7
--- /dev/null
+++ b/crates/yaak-query/src/parser.rs
@@ -0,0 +1,393 @@
+//! Builds the [`Node`] tree from tokens.
+//!
+//! Precedence, tightest first: `-`, then AND (implicit or written), then `OR`. The parser is
+//! forgiving and never fails: a stray `)` or a dangling `OR`, `AND` or `-` is ignored, an
+//! unclosed `(` closes at the end, and an empty group is dropped. Groups and negations nest
+//! at most [`MAX_DEPTH`] deep; past that, `(` and `-` are ignored, so pathological input can't
+//! exhaust the stack (the WebAssembly build has a small one).
+
+use crate::ast::{Node, Span};
+use crate::lexer::{Tok, TokKind, lex};
+
+/// How deep groups and negations may nest, together.
+pub const MAX_DEPTH: usize = 32;
+
+/// Parses a query. `None` when it has no terms (empty, only whitespace, or only operators).
+pub fn parse(input: &str) -> Option<Node> {
+    let toks = lex(input);
+    let mut parser = Parser { toks: &toks, i: 0 };
+    let mut parts = Vec::new();
+    // A `)` with no group to close ends parse_or early; skip it and carry on
+    loop {
+        if let Some(node) = parser.parse_or(0) {
+            parts.push(node);
+        }
+        if parser.peek().is_none() {
+            break;
+        }
+        parser.i += 1;
+    }
+    join(parts, |children, span| Node::And { children, span })
+}
+
+struct Parser<'a> {
+    toks: &'a [Tok],
+    i: usize,
+}
+
+impl<'a> Parser<'a> {
+    fn peek(&self) -> Option<&'a Tok> {
+        self.toks.get(self.i)
+    }
+
+    fn parse_or(&mut self, depth: usize) -> Option<Node> {
+        let mut parts = Vec::new();
+        loop {
+            if let Some(node) = self.parse_and(depth) {
+                parts.push(node);
+            }
+            match self.peek() {
+                Some(Tok { kind: TokKind::Or, .. }) => self.i += 1,
+                _ => break,
+            }
+        }
+        join(parts, |children, span| Node::Or { children, span })
+    }
+
+    fn parse_and(&mut self, depth: usize) -> Option<Node> {
+        let mut parts = Vec::new();
+        while let Some(tok) = self.peek() {
+            match tok.kind {
+                TokKind::Or | TokKind::RParen => break,
+                TokKind::And => self.i += 1,
+                _ => {
+                    if let Some(node) = self.parse_unary(depth) {
+                        parts.push(node);
+                    }
+                }
+            }
+        }
+        join(parts, |children, span| Node::And { children, span })
+    }
+
+    /// Always consumes at least one token.
+    fn parse_unary(&mut self, depth: usize) -> Option<Node> {
+        let tok = self.peek()?;
+        if tok.kind != TokKind::Minus {
+            return self.parse_primary(depth);
+        }
+        self.i += 1;
+        // `-` negates only a term, group or another `-` written directly after it
+        let next = self.peek().filter(|next| {
+            next.span.start == tok.span.end
+                && !matches!(next.kind, TokKind::RParen | TokKind::Or | TokKind::And)
+        });
+        if next.is_none() || depth >= MAX_DEPTH {
+            return None;
+        }
+        let node = self.parse_unary(depth + 1)?;
+        let span = tok.span.to(node.span());
+        Some(Node::Not { node: Box::new(node), span })
+    }
+
+    /// Always consumes at least one token.
+    fn parse_primary(&mut self, depth: usize) -> Option<Node> {
+        let tok = self.peek()?;
+        self.i += 1;
+        match &tok.kind {
+            TokKind::Word(text) => Some(Node::Term { text: text.clone(), span: tok.span }),
+            TokKind::Phrase { text } => Some(Node::Phrase { text: text.clone(), span: tok.span }),
+            TokKind::Field { path, path_span, value, value_span } => Some(Node::Field {
+                path: path.clone(),
+                value: value.clone(),
+                span: tok.span,
+                path_span: *path_span,
+                value_span: *value_span,
+            }),
+            // Too deep: the group's contents carry on at this level instead
+            TokKind::LParen if depth >= MAX_DEPTH => None,
+            TokKind::LParen => {
+                let inner = self.parse_or(depth + 1);
+                if let Some(Tok { kind: TokKind::RParen, .. }) = self.peek() {
+                    self.i += 1;
+                }
+                inner
+            }
+            // Operators are handled by the callers; reaching one here means it is out of place
+            TokKind::Minus | TokKind::Or | TokKind::And | TokKind::RParen => None,
+        }
+    }
+}
+
+/// One node stays itself; several are joined by `make` with a span covering them all.
+fn join(mut parts: Vec<Node>, make: impl Fn(Vec<Node>, Span) -> Node) -> Option<Node> {
+    match parts.len() {
+        0 => None,
+        1 => parts.pop(),
+        _ => {
+            let span = parts[0].span().to(parts[parts.len() - 1].span());
+            Some(make(parts, span))
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::ast::{CompareOp, FieldValue};
+
+    /// A compact rendering of the tree, to keep the tests readable.
+    fn show(input: &str) -> String {
+        fn go(node: &Node) -> String {
+            match node {
+                Node::And { children, .. } => {
+                    format!("(and {})", children.iter().map(go).collect::<Vec<_>>().join(" "))
+                }
+                Node::Or { children, .. } => {
+                    format!("(or {})", children.iter().map(go).collect::<Vec<_>>().join(" "))
+                }
+                Node::Not { node, .. } => format!("(not {})", go(node)),
+                Node::Term { text, .. } => text.clone(),
+                Node::Phrase { text, .. } => format!("{text:?}"),
+                Node::Field { path, value, .. } => {
+                    let value = match value {
+                        FieldValue::Text { text, quoted: false } => text.clone(),
+                        FieldValue::Text { text, quoted: true } => format!("{text:?}"),
+                        FieldValue::Wildcard { pattern } => format!("wild({pattern})"),
+                        FieldValue::Exists => "exists".to_string(),
+                        FieldValue::Compare { op, number } => {
+                            format!("{}{}", op.as_str(), number.raw)
+                        }
+                        FieldValue::Range { low, high } => format!("[{} TO {}]", low.raw, high.raw),
+                    };
+                    format!("@{}={}", path.join("."), value)
+                }
+            }
+        }
+        parse(input).map_or("none".to_string(), |n| go(&n))
+    }
+
+    #[test]
+    fn empty() {
+        assert_eq!(show(""), "none");
+        assert_eq!(show("   "), "none");
+        assert_eq!(show("OR AND"), "none");
+        assert_eq!(show("()"), "none");
+        assert_eq!(show("-"), "none");
+        assert_eq!(show(")"), "none");
+    }
+
+    #[test]
+    fn terms_and_phrases() {
+        assert_eq!(show("foo"), "foo");
+        assert_eq!(show("foo bar"), "(and foo bar)");
+        assert_eq!(show("\"hi there\""), "\"hi there\"");
+        assert_eq!(show("\"hi there"), "\"hi there\"");
+        assert_eq!(show("a\"b c\"d"), "(and a \"b c\" d)");
+        assert_eq!(show(r#""say \"hi\"""#), r#""say \"hi\"""#);
+    }
+
+    #[test]
+    fn boolean_operators() {
+        assert_eq!(show("a OR b"), "(or a b)");
+        assert_eq!(show("a AND b"), "(and a b)");
+        assert_eq!(show("a b OR c"), "(or (and a b) c)");
+        assert_eq!(show("a OR b c"), "(or a (and b c))");
+        assert_eq!(show("a OR b OR c"), "(or a b c)");
+        // Case-sensitive, as Datadog: lowercase is a word
+        assert_eq!(show("a or b"), "(and a or b)");
+        assert_eq!(show("a and b"), "(and a and b)");
+        // NOT is not a keyword
+        assert_eq!(show("NOT a"), "(and NOT a)");
+    }
+
+    #[test]
+    fn dangling_operators_are_ignored() {
+        assert_eq!(show("a OR"), "a");
+        assert_eq!(show("OR a"), "a");
+        assert_eq!(show("a OR OR b"), "(or a b)");
+        assert_eq!(show("a AND"), "a");
+        assert_eq!(show("a -"), "a");
+        assert_eq!(show("- a"), "a");
+        assert_eq!(show("(a -)"), "a");
+    }
+
+    #[test]
+    fn groups() {
+        assert_eq!(show("(a OR b) -c"), "(and (or a b) (not c))");
+        assert_eq!(show("-(a OR b)"), "(not (or a b))");
+        assert_eq!(show("(a OR b"), "(or a b)");
+        assert_eq!(show("a) b"), "(and a b)");
+        assert_eq!(show("((a))"), "a");
+        assert_eq!(show("x (a (b OR c))"), "(and x (and a (or b c)))");
+    }
+
+    #[test]
+    fn negation() {
+        assert_eq!(show("-foo"), "(not foo)");
+        assert_eq!(show("--foo"), "(not (not foo))");
+        assert_eq!(show("-\"a b\""), "(not \"a b\")");
+        assert_eq!(show("-@type:grpc"), "(not @type=grpc)");
+        assert_eq!(show("a-b"), "a-b");
+        assert_eq!(show("-OR a"), "a");
+    }
+
+    #[test]
+    fn fields() {
+        assert_eq!(show("@method:post"), "@method=post");
+        assert_eq!(show("@url:\"a b\""), "@url=\"a b\"");
+        assert_eq!(show("@header.x-cache:HIT"), "@header.x-cache=HIT");
+        assert_eq!(show("@req.header.content-type:json"), "@req.header.content-type=json");
+        assert_eq!(show("@folder:x OR @folder:y"), "(or @folder=x @folder=y)");
+        assert_eq!(show("@url:https://a.b/c?d=e"), "@url=https://a.b/c?d=e");
+    }
+
+    #[test]
+    fn at_is_required() {
+        assert_eq!(show("method:post"), "method:post");
+        assert_eq!(show("localhost:3000"), "localhost:3000");
+        assert_eq!(show("@method"), "@method");
+        assert_eq!(show("@"), "@");
+        assert_eq!(show("@.a:b"), "@.a:b");
+        assert_eq!(show("@a.:b"), "@a.:b");
+        assert_eq!(show("@a b:c"), "(and @a b:c)");
+    }
+
+    #[test]
+    fn trailing_field_has_empty_value() {
+        assert_eq!(show("@method:"), "@method=");
+        assert_eq!(show("@method: post"), "(and @method= post)");
+        assert_eq!(show("(@method:)"), "@method=");
+    }
+
+    #[test]
+    fn field_values() {
+        assert_eq!(show("@url:-foo"), "@url=-foo");
+        assert_eq!(show("@url:AND"), "@url=AND");
+        assert_eq!(show("@url:a(b"), "@url=a(b");
+        assert_eq!(show("(@url:a OR @url:b)"), "(or @url=a @url=b)");
+        assert_eq!(show("@url:a\"b"), "@url=a\"b");
+        assert_eq!(show("@url:\"a"), "@url=\"a\"");
+    }
+
+    #[test]
+    fn special_values() {
+        assert_eq!(show("@status:*"), "@status=exists");
+        assert_eq!(show("-@status:*"), "(not @status=exists)");
+        assert_eq!(show("@url:*.github.com"), "@url=wild(*.github.com)");
+        assert_eq!(show("@url:\"*.github.com\""), "@url=\"*.github.com\"");
+        assert_eq!(show("@status:>=400"), "@status=>=400");
+        assert_eq!(show("@duration:>500ms"), "@duration=>500ms");
+        assert_eq!(show("@status:<1"), "@status=<1");
+        assert_eq!(show("@status:[400 TO 499]"), "@status=[400 TO 499]");
+        assert_eq!(show("(@status:[400 TO 499])"), "@status=[400 TO 499]");
+        // Not quite the shape: literal text or a wildcard
+        assert_eq!(show("@status:>="), "@status=>=");
+        assert_eq!(show("@status:>abc"), "@status=>abc");
+        assert_eq!(show("@status:**"), "@status=wild(**)");
+        assert_eq!(show("@status:[400"), "@status=[400");
+    }
+
+    #[test]
+    fn templates_parse_sanely() {
+        assert_eq!(show("@url:${[ base_url ]}"), "(and @url=${[ base_url ]})");
+        assert_eq!(show("@url:[ base_url ]"), "(and @url=[ base_url ])");
+        assert_eq!(show("@url:\"${[ base_url ]}/x\""), "@url=\"${[ base_url ]}/x\"");
+    }
+
+    #[test]
+    fn compare_value_details() {
+        let Some(Node::Field { value: FieldValue::Compare { op, number }, .. }) =
+            parse("@duration:<=1.5s")
+        else {
+            panic!("not a comparison");
+        };
+        assert_eq!(op, CompareOp::Lte);
+        assert_eq!(number.value, 1.5);
+        assert_eq!(number.unit.as_deref(), Some("s"));
+    }
+
+    #[test]
+    fn spans() {
+        let input = "-@url:\"a b\" OR (c";
+        let node = parse(input).unwrap();
+        let Node::Or { children, span } = &node else {
+            panic!()
+        };
+        assert_eq!(span.slice(input), input);
+        let Node::Not { node: field, span } = &children[0] else {
+            panic!()
+        };
+        assert_eq!(span.slice(input), "-@url:\"a b\"");
+        let Node::Field { span, path_span, value_span, .. } = field.as_ref() else {
+            panic!()
+        };
+        assert_eq!(span.slice(input), "@url:\"a b\"");
+        assert_eq!(path_span.slice(input), "@url:");
+        assert_eq!(value_span.slice(input), "\"a b\"");
+        assert_eq!(children[1].span().slice(input), "c");
+
+        let input = "@method:";
+        let Some(Node::Field { value_span, .. }) = parse(input) else {
+            panic!()
+        };
+        assert_eq!(value_span, Span::new(8, 8));
+    }
+
+    #[test]
+    fn nesting_is_bounded() {
+        fn depth(node: &Node) -> usize {
+            match node {
+                Node::And { children, .. } | Node::Or { children, .. } => {
+                    children.iter().map(depth).max().unwrap_or(0) + 1
+                }
+                Node::Not { node, .. } => depth(node) + 1,
+                _ => 1,
+            }
+        }
+        let parens = format!("{}a OR b{}", "(".repeat(5000), ")".repeat(5000));
+        let negations = format!("{}a", "-".repeat(5000));
+        let mixed = format!("{}a{}", "-(".repeat(5000), ")".repeat(5000));
+        for input in [parens, negations, mixed] {
+            let node = parse(&input).unwrap();
+            assert!(depth(&node) <= MAX_DEPTH + 2, "too deep for {:?}", &input[..10]);
+        }
+        // Within the limit, nesting is kept exactly
+        let input = format!("{}a{}", "(".repeat(MAX_DEPTH), ")".repeat(MAX_DEPTH));
+        assert_eq!(show(&input), "a");
+        assert_eq!(show(&format!("{}a", "-".repeat(2))), "(not (not a))");
+    }
+
+    #[test]
+    fn never_panics_on_odd_input() {
+        for input in [
+            "\"",
+            "\\",
+            "@",
+            "@:",
+            "@a:",
+            "-",
+            "--",
+            "(",
+            ")",
+            "((",
+            "))",
+            "[",
+            "@a:[",
+            "@a:[1 TO",
+            "@a:\"",
+            "@a:\"\\",
+            "é@ü:ö",
+            "@a.b.c:",
+            "@a:>",
+            "@a:<=",
+            "-(",
+            "OR",
+            "AND OR",
+            "@a:)",
+            "🦀 @x:🦀 \"🦀",
+        ] {
+            let _ = parse(input);
+        }
+    }
+}
diff --git a/crates/yaak-query/src/schema.rs b/crates/yaak-query/src/schema.rs
new file mode 100644
index 000000000..17a1c02fb
--- /dev/null
+++ b/crates/yaak-query/src/schema.rs
@@ -0,0 +1,307 @@
+//! What a consumer's fields are, and how its bare words match.
+
+use crate::ast::{FieldValue, Node, NumberLiteral, Span};
+use crate::lexer::parse_number;
+use serde::{Deserialize, Serialize};
+
+/// A consumer's searchable fields.
+#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub struct Schema {
+    pub fields: Vec<FieldDef>,
+    /// Substring when left out.
+    #[serde(default)]
+    #[cfg_attr(test, ts(optional, as = "Option<FreeText>"))]
+    pub free_text: FreeText,
+}
+
+/// How a bare word matches the document's free text. A quoted phrase is always a substring.
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum FreeText {
+    /// A substring, falling back to word-prefix matching (`mr` finds "My Request"), as
+    /// fuzzbunny does.
+    Fuzzy,
+    /// A case-insensitive substring.
+    #[default]
+    Substring,
+}
+
+/// One field: `@name:value`.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub struct FieldDef {
+    /// The field's path, dotted when nested: `method`, `req.header`.
+    pub name: String,
+    pub kind: FieldKind,
+    /// A map of named values, such as headers: `@header.<key>:value` is valid for any key, and
+    /// the key is matched against the document's keys (case-insensitively).
+    #[serde(default)]
+    #[cfg_attr(test, ts(optional, as = "Option<bool>"))]
+    pub map: bool,
+    /// A short note shown with the field in completions.
+    #[serde(default)]
+    #[cfg_attr(test, ts(optional))]
+    pub info: Option<String>,
+}
+
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "camelCase")]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum FieldKind {
+    Text,
+    /// A number, with the units a comparison or range may be written in. A number written
+    /// without a unit is in the field's base unit.
+    Number {
+        #[serde(default)]
+        units: Vec<Unit>,
+    },
+    /// True or false: `@name:true`, `@name:false`, and `@name:*` for true.
+    Flag,
+}
+
+/// A unit a number may be written in, and how many of the field's base unit it is worth:
+/// `{ name: "s", factor: 1000 }` for a field counted in milliseconds.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub struct Unit {
+    pub name: String,
+    pub factor: f64,
+}
+
+/// A field path resolved against the schema.
+#[derive(Clone, Debug, PartialEq)]
+pub struct ResolvedField<'a> {
+    pub def: &'a FieldDef,
+    /// The key under a map field: `x-cache` for `@header.x-cache`. `None` for the map field
+    /// itself (`@header:...`, which matches any key) and for plain fields.
+    pub key: Option<String>,
+}
+
+impl ResolvedField<'_> {
+    /// The lowercase dotted path a document stores this field's values under.
+    pub fn document_path(&self) -> String {
+        match &self.key {
+            Some(key) => format!("{}.{}", self.def.name.to_lowercase(), key.to_lowercase()),
+            None => self.def.name.to_lowercase(),
+        }
+    }
+}
+
+/// Something in a query the schema does not accept. The query still parses and runs; the
+/// offending term matches nothing.
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "camelCase")]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum Diagnostic {
+    /// No field has this path. `span` covers `@path:`.
+    UnknownField { path: String, span: Span },
+    /// The value cannot apply to this field: a unit the field does not take, a number for a
+    /// text-only flag, and so on. `span` covers the value.
+    InvalidValue { path: String, span: Span },
+}
+
+impl Schema {
+    pub fn new(fields: Vec<FieldDef>, free_text: FreeText) -> Self {
+        Self { fields, free_text }
+    }
+
+    /// Finds the field a path names, case-insensitively. A path under a map field resolves to
+    /// that field with the rest of the path as the key; the longest matching map wins.
+    pub fn resolve(&self, path: &[String]) -> Option<ResolvedField<'_>> {
+        let joined = path.join(".").to_lowercase();
+        if let Some(def) = self.fields.iter().find(|f| f.name.to_lowercase() == joined) {
+            return Some(ResolvedField { def, key: None });
+        }
+        self.fields
+            .iter()
+            .filter(|f| f.map)
+            .filter_map(|f| {
+                let name = f.name.to_lowercase();
+                let key = joined.strip_prefix(&name)?.strip_prefix('.')?;
+                (!key.is_empty()).then_some((name.len(), f, key))
+            })
+            .max_by_key(|(len, ..)| *len)
+            .map(|(_, def, key)| ResolvedField { def, key: Some(key.to_string()) })
+    }
+
+    /// The terms in a query this schema cannot run, for the UI to mark.
+    pub fn diagnostics(&self, root: &Node) -> Vec<Diagnostic> {
+        let mut out = Vec::new();
+        root.walk(&mut |node| {
+            let Node::Field { path, value, path_span, value_span, .. } = node else {
+                return;
+            };
+            let joined = path.join(".");
+            match self.resolve(path) {
+                None => out.push(Diagnostic::UnknownField { path: joined, span: *path_span }),
+                Some(field) => {
+                    if !value_fits(&field.def.kind, value) {
+                        out.push(Diagnostic::InvalidValue { path: joined, span: *value_span });
+                    }
+                }
+            }
+        });
+        out
+    }
+}
+
+/// Whether a value can mean anything for a field of this kind.
+fn value_fits(kind: &FieldKind, value: &FieldValue) -> bool {
+    match (kind, value) {
+        (FieldKind::Number { units }, FieldValue::Compare { number, .. }) => {
+            to_base(number, units).is_some()
+        }
+        (FieldKind::Number { units }, FieldValue::Range { low, high }) => {
+            to_base(low, units).is_some() && to_base(high, units).is_some()
+        }
+        // A bare number is equality, so its unit has to be one the field takes
+        (FieldKind::Number { units }, FieldValue::Text { text, quoted: false }) => {
+            parse_number(text).is_none_or(|number| to_base(&number, units).is_some())
+        }
+        // Text holding numbers compares them as plain numbers, without units
+        (FieldKind::Text, FieldValue::Compare { number, .. }) => number.unit.is_none(),
+        (FieldKind::Text, FieldValue::Range { low, high }) => {
+            low.unit.is_none() && high.unit.is_none()
+        }
+        (FieldKind::Flag, FieldValue::Text { text, .. }) => {
+            text.is_empty() || parse_flag(text).is_some()
+        }
+        (FieldKind::Flag, FieldValue::Exists) => true,
+        (FieldKind::Flag, _) => false,
+        _ => true,
+    }
+}
+
+/// A number in the field's base unit, or `None` when its unit is not one the field takes.
+pub(crate) fn to_base(number: &NumberLiteral, units: &[Unit]) -> Option<f64> {
+    match &number.unit {
+        None => Some(number.value),
+        Some(unit) => units
+            .iter()
+            .find(|u| u.name.eq_ignore_ascii_case(unit))
+            .map(|u| number.value * u.factor),
+    }
+}
+
+pub(crate) fn parse_flag(text: &str) -> Option<bool> {
+    match text.to_lowercase().as_str() {
+        "true" | "yes" | "1" => Some(true),
+        "false" | "no" | "0" => Some(false),
+        _ => None,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::parse;
+
+    fn field(name: &str, kind: FieldKind, map: bool) -> FieldDef {
+        FieldDef { name: name.to_string(), kind, map, info: None }
+    }
+
+    fn schema() -> Schema {
+        Schema::new(
+            vec![
+                field("url", FieldKind::Text, false),
+                field("header", FieldKind::Text, true),
+                field("req.header", FieldKind::Text, true),
+                field(
+                    "duration",
+                    FieldKind::Number { units: vec![Unit { name: "s".into(), factor: 1000.0 }] },
+                    false,
+                ),
+                field("error", FieldKind::Flag, false),
+            ],
+            FreeText::Substring,
+        )
+    }
+
+    fn path(p: &str) -> Vec<String> {
+        p.split('.').map(String::from).collect()
+    }
+
+    #[test]
+    fn resolves_plain_and_map_fields() {
+        let s = schema();
+        assert_eq!(s.resolve(&path("URL")).unwrap().def.name, "url");
+        assert_eq!(s.resolve(&path("url")).unwrap().key, None);
+
+        let h = s.resolve(&path("header.X-Cache")).unwrap();
+        assert_eq!(h.def.name, "header");
+        assert_eq!(h.key.as_deref(), Some("x-cache"));
+        assert_eq!(h.document_path(), "header.x-cache");
+
+        let r = s.resolve(&path("req.header.accept")).unwrap();
+        assert_eq!(r.def.name, "req.header");
+        assert_eq!(r.key.as_deref(), Some("accept"));
+
+        // The map itself, with no key
+        assert_eq!(s.resolve(&path("header")).unwrap().key, None);
+
+        assert!(s.resolve(&path("nope")).is_none());
+        assert!(s.resolve(&path("url.x")).is_none());
+        assert!(s.resolve(&path("req")).is_none());
+    }
+
+    #[test]
+    fn reports_unknown_fields_and_bad_values() {
+        let s = schema();
+        let input = "@nope:x @url:y @duration:>5m @duration:>5s @error:maybe @error:true";
+        let diagnostics = s.diagnostics(&parse(input).unwrap());
+        assert_eq!(
+            diagnostics,
+            vec![
+                Diagnostic::UnknownField { path: "nope".into(), span: Span::new(0, 6) },
+                Diagnostic::InvalidValue { path: "duration".into(), span: Span::new(25, 28) },
+                Diagnostic::InvalidValue { path: "error".into(), span: Span::new(50, 55) },
+            ]
+        );
+    }
+
+    /// Diagnostics flag exactly the unit mistakes the evaluator rejects.
+    #[test]
+    fn reports_units_the_evaluator_rejects() {
+        let s = schema();
+        let invalid = |input: &str| !s.diagnostics(&parse(input).unwrap()).is_empty();
+        assert!(invalid("@duration:1m"));
+        assert!(invalid("@duration:[1m TO 2s]"));
+        assert!(invalid("@url:>1s"));
+        assert!(invalid("@url:[1 TO 2s]"));
+        assert!(!invalid("@duration:1s"));
+        assert!(!invalid("@duration:1500"));
+        assert!(!invalid("@duration:abc"));
+        assert!(!invalid("@duration:\"1m\""));
+        assert!(!invalid("@url:>1"));
+        assert!(!invalid("@url:1m"));
+    }
+
+    #[test]
+    fn deserializes_from_json() {
+        let s: Schema = serde_json::from_str(
+            r#"{
+                "fields": [
+                    { "name": "method", "kind": { "type": "text" } },
+                    { "name": "header", "kind": { "type": "text" }, "map": true },
+                    { "name": "size", "kind": { "type": "number", "units": [{ "name": "kb", "factor": 1024 }] } },
+                    { "name": "error", "kind": { "type": "flag" }, "info": "Failed" }
+                ],
+                "freeText": "fuzzy"
+            }"#,
+        )
+        .unwrap();
+        assert_eq!(s.free_text, FreeText::Fuzzy);
+        assert!(s.fields[1].map);
+        assert_eq!(s.fields[3].info.as_deref(), Some("Failed"));
+    }
+}
diff --git a/crates/yaak-query/src/tokens.rs b/crates/yaak-query/src/tokens.rs
new file mode 100644
index 000000000..11ff84243
--- /dev/null
+++ b/crates/yaak-query/src/tokens.rs
@@ -0,0 +1,100 @@
+//! Spans for syntax highlighting.
+
+use crate::ast::Span;
+use crate::lexer::{TokKind, lex};
+use serde::{Deserialize, Serialize};
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub enum TokenKind {
+    /// `@path:`, through the colon.
+    Field,
+    /// A field's value, quoted or not.
+    Value,
+    /// `OR` and `AND`.
+    Operator,
+    /// A quoted phrase in free text.
+    Quote,
+    Paren,
+    /// A `-` that negates what follows it.
+    Negation,
+    /// A bare free-text word.
+    Text,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[cfg_attr(test, derive(ts_rs::TS))]
+#[cfg_attr(test, ts(export, export_to = "query.ts"))]
+pub struct Token {
+    pub kind: TokenKind,
+    pub span: Span,
+}
+
+/// The highlighted pieces of `input`, in order. Whitespace is not covered.
+pub fn tokens(input: &str) -> Vec<Token> {
+    let mut out = Vec::new();
+    let toks = lex(input);
+    for (i, tok) in toks.iter().enumerate() {
+        let kind = match &tok.kind {
+            TokKind::LParen | TokKind::RParen => TokenKind::Paren,
+            TokKind::Minus => {
+                // The parser ignores a `-` with nothing to negate directly after it
+                let negates = toks.get(i + 1).is_some_and(|next| {
+                    next.span.start == tok.span.end
+                        && !matches!(next.kind, TokKind::RParen | TokKind::Or | TokKind::And)
+                });
+                if negates { TokenKind::Negation } else { TokenKind::Text }
+            }
+            TokKind::Or | TokKind::And => TokenKind::Operator,
+            TokKind::Word(_) => TokenKind::Text,
+            TokKind::Phrase { .. } => TokenKind::Quote,
+            TokKind::Field { path_span, value_span, .. } => {
+                out.push(Token { kind: TokenKind::Field, span: *path_span });
+                if value_span.start != value_span.end {
+                    out.push(Token { kind: TokenKind::Value, span: *value_span });
+                }
+                continue;
+            }
+        };
+        out.push(Token { kind, span: tok.span });
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn show(input: &str) -> Vec<(TokenKind, &str)> {
+        tokens(input).into_iter().map(|t| (t.kind, t.span.slice(input))).collect()
+    }
+
+    #[test]
+    fn kinds_and_spans() {
+        use TokenKind::*;
+        assert_eq!(
+            show("-@method:post OR (\"a b\" c) @url:"),
+            vec![
+                (Negation, "-"),
+                (Field, "@method:"),
+                (Value, "post"),
+                (Operator, "OR"),
+                (Paren, "("),
+                (Quote, "\"a b\""),
+                (Text, "c"),
+                (Paren, ")"),
+                (Field, "@url:"),
+            ]
+        );
+        assert_eq!(
+            show("@status:[400 TO 499]"),
+            vec![(Field, "@status:"), (Value, "[400 TO 499]")]
+        );
+        assert_eq!(show("localhost:3000"), vec![(Text, "localhost:3000")]);
+        assert_eq!(show(""), vec![]);
+        // A `-` with nothing directly after it negates nothing
+        assert_eq!(show("- a -"), vec![(Text, "-"), (Text, "a"), (Text, "-")]);
+    }
+}
diff --git a/crates/yaak-query/src/wasm.rs b/crates/yaak-query/src/wasm.rs
new file mode 100644
index 000000000..fbf0089b3
--- /dev/null
+++ b/crates/yaak-query/src/wasm.rs
@@ -0,0 +1,161 @@
+//! The `@yaakapp-internal/query` exports.
+//!
+//! JavaScript counts string positions in UTF-16 code units, so every span going out is
+//! converted from bytes, and the cursor coming in is converted to bytes.
+
+use crate::{Document, Matcher, Schema, Span, Table};
+use serde::Serialize;
+use serde::de::DeserializeOwned;
+use std::collections::HashMap;
+use wasm_bindgen::JsValue;
+use wasm_bindgen::prelude::wasm_bindgen;
+
+/// Byte offset to UTF-16 offset, and back, for one input.
+struct Offsets {
+    /// `utf16[i]` is the UTF-16 offset of byte `i`, for every char boundary and the end.
+    utf16: Vec<u32>,
+}
+
+impl Offsets {
+    fn new(input: &str) -> Self {
+        let mut utf16 = vec![0; input.len() + 1];
+        let mut units = 0u32;
+        for (i, c) in input.char_indices() {
+            for b in 0..c.len_utf8() {
+                utf16[i + b] = units;
+            }
+            units += c.len_utf16() as u32;
+        }
+        utf16[input.len()] = units;
+        Self { utf16 }
+    }
+
+    fn span(&self, span: Span) -> Span {
+        Span { start: self.utf16[span.start()], end: self.utf16[span.end()] }
+    }
+
+    /// The byte offset of UTF-16 offset `pos`, or of the char it falls inside.
+    fn byte(&self, pos: u32) -> usize {
+        self.utf16.partition_point(|&u| u < pos).min(self.utf16.len() - 1)
+    }
+}
+
+fn to_js<T: Serialize>(value: &T) -> Result<JsValue, JsValue> {
+    value
+        .serialize(&serde_wasm_bindgen::Serializer::json_compatible())
+        .map_err(|e| JsValue::from_str(&e.to_string()))
+}
+
+fn from_js<T: DeserializeOwned>(value: JsValue) -> Result<T, JsValue> {
+    serde_wasm_bindgen::from_value(value).map_err(|e| JsValue::from_str(&e.to_string()))
+}
+
+/// The parsed tree (`Node | null`), with UTF-16 spans.
+#[wasm_bindgen]
+pub fn parse_query(input: &str) -> Result<JsValue, JsValue> {
+    let offsets = Offsets::new(input);
+    let mut root = crate::parse(input);
+    if let Some(root) = &mut root {
+        root.map_spans(&|s| offsets.span(s));
+    }
+    to_js(&root)
+}
+
+/// Highlighting spans (`Token[]`), in UTF-16 offsets.
+#[wasm_bindgen]
+pub fn query_tokens(input: &str) -> Result<JsValue, JsValue> {
+    let offsets = Offsets::new(input);
+    let mut tokens = crate::tokens(input);
+    for token in &mut tokens {
+        token.span = offsets.span(token.span);
+    }
+    to_js(&tokens)
+}
+
+/// `Completions | null` at UTF-16 offset `cursor`. `values` maps field paths to the values seen
+/// for them (`Record<string, string[]>`).
+#[wasm_bindgen]
+pub fn query_completions(
+    input: &str,
+    cursor: u32,
+    schema: JsValue,
+    values: JsValue,
+) -> Result<JsValue, JsValue> {
+    let schema: Schema = from_js(schema)?;
+    let values: HashMap<String, Vec<String>> =
+        if values.is_undefined() || values.is_null() { HashMap::new() } else { from_js(values)? };
+    let offsets = Offsets::new(input);
+    let mut result = crate::completions(input, offsets.byte(cursor), &schema, &values);
+    if let Some(result) = &mut result {
+        result.span = offsets.span(result.span);
+    }
+    to_js(&result)
+}
+
+/// The terms the schema cannot run (`Diagnostic[]`), with UTF-16 spans.
+#[wasm_bindgen]
+pub fn query_diagnostics(input: &str, schema: JsValue) -> Result<JsValue, JsValue> {
+    let schema: Schema = from_js(schema)?;
+    let offsets = Offsets::new(input);
+    let mut diagnostics =
+        crate::parse(input).map(|root| schema.diagnostics(&root)).unwrap_or_default();
+    for diagnostic in &mut diagnostics {
+        match diagnostic {
+            crate::Diagnostic::UnknownField { span, .. }
+            | crate::Diagnostic::InvalidValue { span, .. } => *span = offsets.span(*span),
+        }
+    }
+    to_js(&diagnostics)
+}
+
+#[wasm_bindgen]
+pub fn query_format_field(field: &str, value: &str, negated: bool) -> String {
+    crate::format_field(field, value, negated)
+}
+
+/// A query parsed and resolved once, then matched against many documents.
+#[wasm_bindgen]
+pub struct QueryMatcher {
+    matcher: Matcher,
+}
+
+#[wasm_bindgen]
+impl QueryMatcher {
+    #[wasm_bindgen(constructor)]
+    pub fn new(input: &str, schema: JsValue) -> Result<QueryMatcher, JsValue> {
+        let schema: Schema = from_js(schema)?;
+        Ok(QueryMatcher { matcher: Matcher::new(crate::parse(input).as_ref(), &schema) })
+    }
+
+    /// Whether the query has no terms, so it matches everything.
+    #[wasm_bindgen(getter, js_name = isEmpty)]
+    pub fn is_empty(&self) -> bool {
+        self.matcher.is_empty()
+    }
+
+    /// Whether a `Document` (`{ text, fields }`), as JSON, matches. JSON crosses into wasm far
+    /// faster than walking a JS object property by property.
+    pub fn matches(&self, doc_json: &str) -> Result<bool, JsValue> {
+        let doc: Document = serde_json::from_str(doc_json).map_err(js_error)?;
+        Ok(self.matcher.matches(&doc))
+    }
+
+    /// The indexes of the matching rows, for many documents in one call. The cells (per row,
+    /// its free text, then one per column) are packed as in [`Table::from_packed`]: joined into
+    /// one string, with each one's UTF-16 length, or `Table::NO_VALUE` for no value.
+    #[wasm_bindgen(js_name = matchingRows)]
+    pub fn matching_rows(
+        &self,
+        columns: Vec<String>,
+        joined: &str,
+        lengths: &[u32],
+    ) -> Result<Vec<u32>, JsValue> {
+        let table = Table::from_packed(columns, joined, lengths)
+            .ok_or_else(|| JsValue::from_str("cells do not line up with the columns"))?;
+        Ok(self.matcher.matching_rows(&table))
+    }
+}
+
+fn js_error(e: impl std::fmt::Display) -> JsValue {
+    JsValue::from_str(&e.to_string())
+}
diff --git a/package-lock.json b/package-lock.json
index a4539a7a3..9a34d7914 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -64,6 +64,7 @@
         "crates/yaak-git",
         "crates/yaak-models",
         "crates/yaak-plugins",
+        "crates/yaak-query",
         "crates/yaak-sse",
         "crates/yaak-sync",
         "crates/yaak-templates",
@@ -322,6 +323,13 @@
       "name": "@yaakapp-internal/plugins",
       "version": "1.0.0"
     },
+    "crates/yaak-query": {
+      "name": "@yaakapp-internal/query",
+      "version": "1.0.0",
+      "devDependencies": {
+        "rimraf": "^6.1.2"
+      }
+    },
     "crates/yaak-sse": {
       "name": "@yaakapp-internal/sse",
       "version": "1.0.0",
@@ -5241,6 +5249,10 @@
       "resolved": "crates-proxy/yaak-proxy-lib",
       "link": true
     },
+    "node_modules/@yaakapp-internal/query": {
+      "resolved": "crates/yaak-query",
+      "link": true
+    },
     "node_modules/@yaakapp-internal/rpc-schema": {
       "resolved": "crates/common/yaak-rpc-schema",
       "link": true
diff --git a/package.json b/package.json
index 9a83a9673..9361af88a 100644
--- a/package.json
+++ b/package.json
@@ -63,6 +63,7 @@
     "crates/yaak-git",
     "crates/yaak-models",
     "crates/yaak-plugins",
+    "crates/yaak-query",
     "crates/yaak-sse",
     "crates/yaak-sync",
     "crates/yaak-templates",
diff --git a/vite.config.ts b/vite.config.ts
index d009150e4..88ca8c9af 100644
--- a/vite.config.ts
+++ b/vite.config.ts
@@ -10,6 +10,7 @@ export default defineConfig({
     ignorePatterns: [
       "**/bindings/**",
       "**/routeTree.gen.ts",
+      "crates/yaak-query/pkg/**",
       "crates/yaak-templates/pkg/**",
       "crates/yaak-wasm/pkg/**",
       "plugins/importer-bruno/tests/fixtures/upstream/**",
@@ -18,6 +19,7 @@ export default defineConfig({
   lint: {
     ignorePatterns: [
       "npm/**",
+      "crates/yaak-query/pkg/**",
       "crates/yaak-templates/pkg/**",
       "crates/yaak-wasm/pkg/**",
       "**/bindings/gen_*.ts",