feat(retrieval): Wilson prior, exploration slot, embeddings + three-leg fusion#83
Merged
Conversation
…tings) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… 48-session gate Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…eted Replaces the popularity + overlooked-weight + ignored-demotion CASE ORDER BY in retrieve_reflections with wilson_lower_bound over (useful+overlooked)/rated, computed in Python and sorted (wilson DESC, confidence DESC, updated_at DESC). Deletes OVERLOOKED_RANKING_WEIGHT, IGNORED_DEMOTION_FLOOR, IGNORED_DEMOTION_WEIGHT from memory_rating.py and their two superseded pinning tests. semantic.py's list_for_project also imported the deleted constants; bridged its ORDER BY to plain popularity (useful_count DESC, created_at DESC) pending Task 3's full Wilson rewrite there. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ries Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…bedder and 60s breaker Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ncEmbedder Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…cEmbedder Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ction edit The management UI built ReflectionService and SemanticMemoryService without a sync_embedder, so UI-driven writes (edit reflection, create/update semantic memory) never populated the vec0 embeddings tables — only MCP/synthesis writes did, requiring a manual CLI backfill. - create_app() gains a sync_embedder kwarg (defaults to auto-detect from get_config().embeddings_backend, mirroring mcp/server.py and storage/sqlite.py); wired into ReflectionService and all SemanticMemoryService construction sites. - ReflectionService.update_text now re-embeds via a new module-level _write_reflection_embedding() (extracted from ReflectionSynthesisService._store_embedding so both classes share it) — previously an edit left a stale vector pointing at the old text. - tests/ui/conftest.py's client fixture pins BETTER_MEMORY_EMBEDDINGS_BACKEND=sqlite so ordinary route tests don't depend on a live Ollama instance (default backend is ollama); tests proving the wiring build their own app with an injected fake embedder. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…self-heal Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ade order Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…contract Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
🔴 Claude BugBot Analysis
Found 1 potential bug in this PR.
medium: 1
Reviewed the retrieval-quality PR (Wilson-score ranking, exploration slot, RRF vector fusion, SyncEmbedder circuit breaker, and embedding write/backfill paths) in depth; the ranking/fusion/bucketing logic all checked out correctly against their tests, but the SyncEmbedder wiring creates two independent circuit-breaker instances instead of the one shared, process-wide breaker the code documents.
…er and backend Addresses PR #83 review: SqliteBackend built a second SyncEmbedder, splitting the process-wide circuit breaker into two independent ones. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
🟢 Claude BugBot Analysis
Reviewed the retrieval-quality PR (Wilson-score ranking, reflection/semantic embeddings, SyncEmbedder circuit breaker, exploration slot) across all changed source files; no genuine new defects found. The previously reported SyncEmbedder-instantiated-twice bug is fixed: build_backend/SqliteBackend now accept and forward the caller's shared sync_embedder instead of constructing a second one.
No bugs were detected in this PR.
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.
Summary
Implements PR-A of the retrieval-quality design (
docs/superpowers/specs/2026-07-23-retrieval-quality-design.md):services/scoring.py) — lower bound (z=1.96) on(useful+overlooked)/(useful+overlooked+ignored), computed in Python; replaces theuseful_count + 3*overlookedORDER BY and the fix(memory): close the rating loop and stop diluting what gets served #81 ignored-demotion CASE (constants deleted). Hit rate now beats raw serve-count: 3/4 outranks 67/192.SyncEmbedder: freshOllamaEmbedder(timeout=5, max_retries=1)per bridge call (httpx client is loop-bound), 60s circuit breaker, never raises. Lazy self-heal on query retrieves (cap 20) + one-shotpython -m better_memory.cli.backfill_embeddings. Migration 0014 addssemantic_embeddings(additive).rrf_k=60matchingsearch/hybrid.py; every leg degrades independently (no embedder / breaker open / no match ⇒ exactly the prior behaviour, test-pinned incl. multi-row order equality).Validation
ignoredat first — the designed cost of acquiring ranking signal; (b) live-installed contextual hooks fired inside A8 sessions (absent in the baseline arm); excluding that channel, z = −1.381. The metric to watch post-merge is liveuseful%trend as exploration converts untested memories into rated ones.Deploy
After merge: restart Claude Code (migration 0014 applies on server start), then
python -m better_memory.cli.backfill_embeddingsto embed the historical corpus.Carried follow-ups (triaged Minor): agentcore backend still uses the legacy linear ranking client-side;
sync_embedfinally-clause exception masking polish;storage/sqlite.pyReflectionService wiring-for-consistency (dead write path today).🤖 Generated with Claude Code