feat(recall): relevance-first seed selection (score-gap) + conditional-trust framing - #78
Open
evilh2019 wants to merge 3 commits into
Open
feat(recall): relevance-first seed selection (score-gap) + conditional-trust framing#78evilh2019 wants to merge 3 commits into
evilh2019 wants to merge 3 commits into
Conversation
…of fixed count - recallPrecise: seeds = nodes above max(top1 - recallScoreGap, recallMinScore), capped by recallSeedCap (default 2x recallMaxNodes); fall back to top-3 when below threshold so community expansion still has seeds - precise path final slice = seed cap (was fixed limit), so the budget goes to relevant nodes only - recallGeneralized: drop time-based communityRepresentatives fallback — it injected the same off-topic 'recent community' nodes every turn regardless of query (measured: 6 fixed noise nodes in every recall) - new GmConfig: recallScoreGap (0.10), recallMinScore (0.58), recallSeedCap - benchmark (14 queries x 7 topics, real bge-m3): coverage 33.5% -> 51.7%, precision 0.37 -> 0.54, tokens 2187 -> 2110 per turn
…fied) Replace the contradictory 'untrusted reference material' framing with a conditional-trust + usage-protocol framing: - recalled memories = pointers/evidence, may be outdated; verify fragile facts (paths/versions/status) before asserting - PATCHES = prefer newer; CONFLICTS_WITH = check conditions - follow skills when trigger conditions match; say so if not covered; never invent specifics absent from the recalled context - assemble.ts: drop 'proven solutions... apply them directly' (contradicts the untrusted line; now aligned with the new framing) A/B (6-question raw-API test, same recall content): factual accuracy unchanged (100% both), source citations 3x (2 vs 6), answers -7% shorter. 18-question expanded run in progress.
…t-match supplement The initial merge applied the relevance-first gap threshold to the fused relevance (semantic + RRF). RRF's 0.35/(rank+1) boost on weak-semantic FTS hits creates artificial spikes: a 0.5-semantic node + 0.35 RRF overtakes the true top (0.72) and raises the threshold, cutting off the genuinely relevant semantic cluster (measured: coverage dropped to 23.9%, several queries 0 hits). Fix: threshold = max(topSemantic - recallScoreGap, recallMinScore, minSemanticScore) applied to PURE semantic scores only; FTS hits join as supplements up to recallSeedCap (preserves upstream's exact-identifier intent). Merged-code benchmark (14 queries x 7 topics, real bge-m3, allowBroadFallback = false like the DSH adapter): coverage 33.5% -> 45.5%, precision 0.37 -> 0.48, tokens 2187 -> 2029/turn.
Author
Production deployment verified ✅This change is now live in the DSH deployment (2026-08-22 17:03 CST):
No rollback needed so far. Benchmark numbers in the PR description were measured against this exact build. |
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
2 participants
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.
Summary
Relevance-first recall: stop cutting the seed set at a fixed count and stop feeding the model contradictory trust guidance. Two independent changes, both measured against the real graph (580 nodes, bge-m3 vectors, DSH deployment):
1. Gap-threshold seed selection (
recaller/recall.ts,types.ts)max(topSemantic - recallScoreGap, recallMinScore, minSemanticScore)) instead of a fixed.slice(0, limit). A fixed count cuts in the middle of a dense score region where ranks 6..12 are nearly identical in relevance — dropping real memories at an arbitrary boundary.recallSeedCap(preserves the hybrid retrieval's exact-identifier intent). RRF is still used for PPR weighting only.recallScoreGap(0.10),recallMinScore(0.58),recallSeedCap(default 2x recallMaxNodes).allowBroadFallback: falsesemantics.2. Conditional-trust framing (
dsh.ts,format/assemble.ts)PATCHESedges, and never invent specifics absent from the recalled context.Measurements
Coverage benchmark (14 queries x 7 topics, real recaller + bge-m3,
allowBroadFallback: false):Prompt A/B (18 questions, identical recall content, V0 vs V1 framing, deepseek-v4-flash): factual accuracy 100% both; V1 cites source memories ~36% more (15 vs 11); answer length unchanged.
Compatibility
allowBroadFallbackdefault is untouched.