feat: configurable search sort (recency/relevance/hybrid) + phrase matching fix#296
Merged
jalehman merged 7 commits intoMartian-Engineering:mainfrom Apr 6, 2026
Conversation
There was a problem hiding this comment.
Pull request overview
This PR improves lcm_grep search quality by adding a configurable sort order (recency/relevance/hybrid) and updating FTS5 query sanitization to preserve user-quoted phrases as phrase queries.
Changes:
- Added
sortparameter tolcm_greptool schema and forwarded it into retrieval. - Implemented
recency/relevance/hybridresult sorting inRetrievalEngine.grep. - Updated FTS5 query sanitization to keep
"..."groups intact as phrases rather than splitting on whitespace.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.
| File | Description |
|---|---|
| src/tools/lcm-grep-tool.ts | Adds sort param to tool schema and forwards it to retrieval. |
| src/retrieval.ts | Introduces sort modes for grep results (recency/relevance/hybrid). |
| src/store/fts5-sanitize.ts | Preserves user-quoted phrases during FTS5 query sanitization. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Contributor
Author
|
Part of the LCM Performance & Cache Optimization Sprint — see #297 for the full tracking issue linking all 5 PRs. |
100yenadmin
pushed a commit
to electricsheephq/lossless-claw-test
that referenced
this pull request
Apr 6, 2026
Addresses PR Martian-Engineering#296 review comments: 1. Hybrid sort math: FTS5 rank is negative (more negative = better). Old formula multiplied negative rank by age factor, making older results rank BETTER (inverted). Fixed: negate rank to get positive relevance score, divide by age penalty, negate result for ascending sort. 2. Sort scope: documented that sort is applied post-LIMIT — the store queries ORDER BY created_at DESC LIMIT ?, so relevance/hybrid modes only re-rank within the top-N-by-recency window. Known limitation. 3. Added 4 phrase matching tests to fts5-sanitize.test.ts: - Multi-word quoted phrase preserved - Multiple quoted phrases preserved - Mixed quoted and unquoted terms - Empty quoted phrase Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
100yenadmin
pushed a commit
to electricsheephq/lossless-claw-test
that referenced
this pull request
Apr 6, 2026
Addresses PR Martian-Engineering#296 review round 3: 1. Relevance sort: added recency tiebreaker when FTS5 ranks are equal, preventing implementation-dependent ordering. 2. Sort mode validation: relevance/hybrid silently fall back to recency when mode is regex (FTS5 rank not available). Previously accepted without effect. 3. Removed unused 'remaining' variable in fts5-sanitize.ts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Contributor
Author
This was referenced Apr 6, 2026
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+265
to
+266
| // AGE_DECAY_RATE: ~1% relevance penalty per 10 hours of age. | ||
| // At 100h: 10% penalty. At 1000h (~6 weeks): 50% penalty. |
There was a problem hiding this comment.
The AGE_DECAY_RATE comment math is slightly off: with agePenalty = 1 + ageHours * 0.001, at 100h the relevance becomes 1/1.1 (~9.1% reduction), not 10%. Either adjust the wording (e.g., “~9%”) or tweak the rate/formula to match the stated percentages so future tuning is based on correct intuition.
Suggested change
| // AGE_DECAY_RATE: ~1% relevance penalty per 10 hours of age. | |
| // At 100h: 10% penalty. At 1000h (~6 weeks): 50% penalty. | |
| // AGE_DECAY_RATE: ~1% increase in the age penalty per 10 hours of age. | |
| // At 100h: ~9% relevance reduction. At 1000h (~6 weeks): 50% reduction. |
Contributor
Author
There was a problem hiding this comment.
Valid — the exact math at 100h is 1/(1+0.1) = 9.1% reduction, not 10%. Will fix the comment to say '~9%' instead of '~10%'. The rate itself is correct — just the human-readable approximation was rounded.
…e matching Two search quality improvements: 1. Add sort parameter to lcm_grep: "recency" (default, preserves current behavior), "relevance" (FTS5 BM25 rank — best match first), or "hybrid" (blends relevance with recency). FTS5 already computes BM25 scores on every search but they were being ignored. The "recency" default preserves the original design intent — the DAG structure already filters for importance through compression, and lcm_expand_query handles smart retrieval via sub-agent LLM reasoning. But for users with 200K+ message histories, relevance sorting surfaces important older matches that recency-only misses. 2. Fix phrase matching in fts5-sanitize.ts: user-quoted phrases like "error handling" were being split into "error" AND "handling" (two separate terms). Now preserved as a single phrase match. Fixes Martian-Engineering#293 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Addresses PR Martian-Engineering#296 review comments: 1. Hybrid sort math: FTS5 rank is negative (more negative = better). Old formula multiplied negative rank by age factor, making older results rank BETTER (inverted). Fixed: negate rank to get positive relevance score, divide by age penalty, negate result for ascending sort. 2. Sort scope: documented that sort is applied post-LIMIT — the store queries ORDER BY created_at DESC LIMIT ?, so relevance/hybrid modes only re-rank within the top-N-by-recency window. Known limitation. 3. Added 4 phrase matching tests to fts5-sanitize.test.ts: - Multi-word quoted phrase preserved - Multiple quoted phrases preserved - Mixed quoted and unquoted terms - Empty quoted phrase Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Addresses PR Martian-Engineering#296 review round 3: 1. Relevance sort: added recency tiebreaker when FTS5 ranks are equal, preventing implementation-dependent ordering. 2. Sort mode validation: relevance/hybrid silently fall back to recency when mode is regex (FTS5 rank not available). Previously accepted without effect. 3. Removed unused 'remaining' variable in fts5-sanitize.ts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ription Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move lcm_grep sort handling into the FTS query layer so relevance and hybrid ordering are applied before LIMIT truncates candidates. Add a changeset, update agent-facing grep guidance and delegated retrieval prompts, and cover the new behavior with focused tests. Regeneration-Prompt: |\n Address the PR review findings on the search-quality work for lcm_grep. Preserve the new sort parameter and quoted-phrase behavior, but fix the implementation so full-text relevance and hybrid ordering happen before LIMIT is applied instead of reranking a recency-truncated window in memory. Keep regex and fallback behavior intact. Also add the missing changeset because this is user-facing plugin behavior. Finally inspect the prompt and tool-instruction surfaces that teach the model how to use LCM recall tools, and update them so agents know when to prefer full_text, when to quote exact phrases, and when relevance or hybrid sorting is appropriate. Add regression tests for the ranking path and prompt guidance.
fff8b3a to
d582a4c
Compare
Contributor
|
Thank you! |
This was referenced Apr 7, 2026
Merged
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
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
Two search quality improvements for `lcm_grep`:
Fixes #293
Why This Matters
FTS5 already computes BM25 relevance scores on every search, but they were being discarded at sort time. On conversations with 200K+ messages, recency-only sorting buries relevant older matches while surfacing recent-but-tangential hits.
Design Rationale (from LCM architecture analysis)
The `"recency"` default is preserved because LCM's design intentionally relies on:
The `"relevance"` option is for users with very long histories where the routing function benefits from surfacing the best keyword match, not just the most recent one. The agent chooses per-query — use recency for "what just happened?" and relevance for "find the discussion about X from last week."
Changes
Test plan