Aska
8495a4b055
feat(memory): hybrid vector search via embedder
...
- repository: optional pgvector embedding column (migration with
CREATE EXTENSION guarded by main), SearchByVector cosine search,
vector-enabled insert path
- service: SetEmbedder — memories are embedded on Add; Search merges
vector hits with n-gram keyword matches (vector first, deduped)
- embedder was already implemented (OpenAI-compatible API + n-gram FNV
fallback) but never instantiated; main now wires it from
EMBEDDING_API_URL/EMBEDDING_MODEL config
2026-08-20 20:57:13 +08:00
Aska
14a8eecfdb
fix(memory): improve dedup thresholds and keyword priority
...
- Dedup: OR logic with 28%/0.45 thresholds (was 25%/0.4)
- Keywords: segments first → 3-grams → 2-grams
- False positive: 10% → acceptable level
- Accuracy: 80% correct dedup, 71% real rate
2026-08-12 19:20:14 +08:00
Aska
dc3671eb8d
feat(bench): add memory benchmark tool + export keywords
...
- tools/memory-bench: 6 metrics (write/search/dedup/keyword/bulk/rate)
- memory: export ExtractKeywordsPublic for benchmarking
- Results: 1000+ 条/秒 write, 380µs search, 80% dedup accuracy, 71% real rate
2026-08-12 19:15:15 +08:00
Aska
886fcf2216
fix(memory): rewrite dedup with n-gram keyword overlap + embedder
...
- Keyword extraction: 2-3 char sliding window + segment splitting
- Dedup: keyword overlap (>25%) OR Jaccard (>0.4) triggers merge
- SearchByNGrams: pgvector GIN index on keywords array
- Embedder: n-gram FNV hash (zero dep), OpenAI API with auto-fallback
- GIN index on keywords column for fast overlap search
- All tests pass
2026-08-12 18:24:04 +08:00
Aska
7295556c59
feat(services): add memory service — model, repository, service layer
...
- services/memory/model: Memory struct with categories and decay
- services/memory/repository: PostgreSQL CRUD + keyword search
- services/memory/service: dedup, decay, keyword extraction
- tests coming in follow-up commit
2026-08-09 20:50:01 +08:00
Aska