starred/minne
1
0
Fork 0
mirror of https://github.com/perstarkse/minne.git synced 2026-07-07 05:15:08 +02:00
Code Issues Packages Projects Releases 10 Wiki Activity

fix: never block fts, rely on rrf

Browse Source
This commit is contained in:
Per Stark
2025-12-22 22:56:57 +01:00
parent 589257e721
commit 97ce77c574
1 changed files with 1 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
retrieval-pipeline/src/pipeline/stages/mod.rs
+1 -21
View File
@@ -572,27 +572,7 @@ pub async fn collect_vector_chunks(ctx: &mut PipelineContext<'_>) -> Result<(),
// .filter(|c| c.scores.vector.is_some() && c.scores.fts.is_some()) // .filter(|c| c.scores.vector.is_some() && c.scores.fts.is_some())
// .count(); // .count();
// If we have very low overlap (few chunks with both signals), drop FTS-only chunks.
// These are often noisy on keyword-heavy datasets and dilute strong vector hits.
// Keep vector-only and “golden” (vector+FTS) chunks.
let fts_only_count = vector_chunks
.iter()
.filter(|c| c.scores.vector.is_none())
.count();
let both_count = vector_chunks
.iter()
.filter(|c| c.scores.vector.is_some() && c.scores.fts.is_some())
.count();
if fts_only_count > 0 && both_count < 3 {
let before_filter = vector_chunks.len();
vector_chunks.retain(|c| c.scores.vector.is_some());
let after_filter = vector_chunks.len();
debug!(
fts_only_filtered = before_filter - after_filter,
both_signals_preserved = both_count,
"Filtered out FTS-only chunks due to low overlap, preserved golden chunks"
);
}
debug!( debug!(
top_fused_scores = ?vector_chunks.iter().take(5).map(|c| c.fused).collect::<Vec<_>>(), top_fused_scores = ?vector_chunks.iter().take(5).map(|c| c.fused).collect::<Vec<_>>(),