Conversation
- P1: Add local embedding provider via fastembed (bge-small-en-v1.5, 384 dims). Runs on-device via ONNX Runtime with no API key needed. Model downloads on first use (~33MB), cached locally. Made 'local' the new default provider; Voyage/OpenAI remain available via .lore.json config. Existing embeddings auto-clear and re-embed on provider change via checkConfigChange(). - P2: Add exact keyword match boost as additional RRF signal in recall. Ranks candidates by verbatim query term overlap to boost proper nouns, file names, and technical terms that BM25 stemming may dilute. - P3: Add regex-based pattern extraction at distillation time. 8 conservative patterns detect decision/preference language in distilled observations and seed knowledge entries via ltm.create(). Runs in both distillSegment() and metaDistill() paths. Zero LLM cost.
…l scoring (#137) ## Summary - **P0: Unified embedding backfill** — Moved embedding backfill from OpenCode-only to a shared `runStartupBackfill()` in core. Both OpenCode and Pi adapters now call this at init, fixing near-zero embedding coverage (~0.5%) for Pi-served projects. Logs coverage stats to stderr at startup for immediate visibility. - **P3: Retroactive metric backfill** — Added `backfillMetrics()` to compute `r_compression` and `c_norm` for ~2,863 pre-v12 distillations by loading source temporal messages. Idempotent, handles pruned sources gracefully. Runs synchronously at startup. - **P2: Quality-weighted recall scoring** — Wired `c_norm` into recall by adding a distillation quality RRF list. Uniform segments (low c_norm) rank higher; old bursty segments rank lower via `c_norm + (ageDays/90) * 0.1`. Mild secondary signal that blends naturally with existing BM25 and vector relevance. ## Files Changed | File | Change | |---|---| | `packages/core/src/embedding.ts` | Added `runStartupBackfill()` with coverage logging | | `packages/core/src/distillation.ts` | Added `backfillMetrics()` for retroactive r_compression/c_norm | | `packages/core/src/recall.ts` | Added c_norm to Distillation type, all SELECTs, and quality RRF list | | `packages/opencode/src/index.ts` | Replaced inline backfill with core functions | | `packages/pi/src/index.ts` | Added embedding + metric backfill calls | | `.lore.md` | Updated project knowledge | ## Testing - 796 tests pass, 0 failures - Build TS errors are pre-existing (fastembed type issue on parent branch)
TypeScript's overload resolution can't infer the correct variant when the model identifier is resolved dynamically from config. Cast the options object to the standard init signature (BGESmallENV15 is the default) to satisfy the type checker. The enum lookup at runtime guarantees a valid model value.
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
Local embedding provider (P1): Adds
fastembed(bge-small-en-v1.5, 384 dims) as default embedding provider. Runs fully on-device via ONNX Runtime — no API key, zero cost, ~150ms per query. Voyage and OpenAI remain available via.lore.jsonconfig. Existing embeddings auto-clear and re-embed on provider change viacheckConfigChange().Exact keyword match boost (P2): Adds
exactTermMatchRank()as an additional RRF signal in recall. Ranks candidates by verbatim query term overlap, boosting proper nouns, file names, and technical terms that BM25's Porter stemming may dilute. Zero-cost, additive-only.Pattern extraction at distillation time (P3): 8 conservative regex patterns detect decision/preference language ("decided to use X", "chose X over Y", "prefers X over Y") in distilled observations and seed knowledge entries via
ltm.create(). Runs in bothdistillSegment()andmetaDistill(). Zero LLM cost, <1ms runtime.Migration
Existing Voyage users with no explicit
.lore.jsonwill silently switch to local embeddings.checkConfigChange()detects the fingerprint mismatch and auto-rebuilds all embeddings with the new provider. To stay on Voyage, add explicit config:{ "search": { "embeddings": { "provider": "voyage", "model": "voyage-code-3", "dimensions": 1024 } } }Test results
796 pass, 0 fail (including 42 new tests across 3 test files).