feat: add RRF score fusion and rewrite recall tool#47
Merged
Conversation
- Add reciprocalRankFusion() to search.ts — merges ranked lists using
RRF (k=60) for cross-source relevance fusion
- Add normalizeRank() for min-max score normalization (display only)
- Add searchScored() to ltm.ts — returns BM25 rank with results
- Add searchScored() to temporal.ts — returns FTS5 rank with results
- Add searchDistillationsScored() to reflect.ts — scored distillation search
- Rewrite recall tool execute() to use scored search + RRF fusion
- All 3 sources searched with scored variants
- Results tagged with source type, fused via RRF
- Single ranked list with source annotations: [knowledge/category],
[distilled], [temporal/role]
- Add 11 new tests for normalizeRank and reciprocalRankFusion
BYK
added a commit
that referenced
this pull request
Mar 22, 2026
## Phase 3 of search improvements (depends on #47) Replaces the coarse bag-of-words term-overlap scoring in `forSession()` with FTS5 BM25-based scoring. ### Problem `forSession()` used manual term-overlap counting: extract top 30 words >3 chars, count how many appear in each entry via `string.includes()`. This ignored: - Porter stemming ("configure" wouldn't match "configuration") - TF-IDF weighting (all matching terms counted equally) - Stopwords (common words inflated match counts) ### Solution **New `scoreEntriesFTS()`** in ltm.ts: - Runs session context terms against `knowledge_fts` using BM25 - Uses **OR** semantics (not AND-then-OR) because we're scoring all candidates for ranking, not searching for exact matches — an entry matching 1 of 40 terms should get a low score, not be excluded - BM25 naturally weights entries matching more terms higher - Scores normalized to 0–1 and multiplied by entry confidence **Improved `extractTopTerms()`** moved to `search.ts`: - Now uses same STOPWORDS set from Phase 1 - Drops single chars only (not >3 char threshold) — preserves "DB", "CI", "IO" - Increased limit from 30 to 40 terms ### Safety net preserved Top 5 project entries by confidence are always included regardless of FTS match, preventing the scoring change from accidentally excluding critical project knowledge. ### Test coverage - 8 new tests for `extractTopTerms()` (stopwords, 2-char tokens, limits, punctuation) - All 12 existing `forSession()` tests continue to pass
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Phase 2 of search improvements (depends on #46)
Adds cross-source score fusion using Reciprocal Rank Fusion and rewrites the recall tool to produce a single ranked result list.
Changes
New in
src/search.tsreciprocalRankFusion<T>()— merges multiple ranked lists using RRF (k=60, Cormack et al. 2009). Rank-based, not score-based, so magnitude differences across FTS tables don't matter.normalizeRank()— min-max normalization of FTS5 BM25 ranks to 0–1 (for display only)New scored search variants
ltm.searchScored()— returnsKnowledgeEntry & { rank }with BM25 scores viabm25(knowledge_fts, 6, 2, 3)temporal.searchScored()— returnsTemporalMessage & { rank }searchDistillationsScored()— returnsDistillation & { rank }All scored variants include AND→OR fallback (same as Phase 1 search functions).
Rewritten recall tool
[knowledge/category],[distilled],[temporal/role])Test coverage
normalizeRank()andreciprocalRankFusion()