v0.5.0 Release Notes

Released 2026-04-18

v0.4.5 -> v0.5.0

Summary

Add semantic vector search backed by OpenAI embeddings. Vectors are inlined on each Entry, persisted through AOF, and queried via new VSEARCH / VSIM / VGET commands; a content-addressed embedding cache under __torii:embed:* de-duplicates API calls across identical texts.

Changes

FEAT

• Add core/openai package — standalone text-embedding-3-small client with Client.Dim / Client.Model accessors, singleton initialization, and .env loading via godotenv
• Add Entry.Vector []float32 field with base64 little-endian float32 codec (core/store/vector.go); AOFRecord.Vector *string is replayed and re-emitted during compaction for lossless restart
• Add embedding cache under __torii:embed:<sha256(model|dim|text)> (core/store/vcache.go); d != currentDim is treated as MISS, and KEYS / FIND / QUERY / VSEARCH skip the __torii:* namespace so internal keys never leak
• Add SET <key> <value> [NX|XX] [<seconds>] VECTOR — main key writes synchronously, a background goroutine attaches the embedding, and Store.Close() joins sync.WaitGroup before compacting AOF
• Add VSEARCH <text> [MATCH <pattern>] [LIMIT <n>] — top-K cosine similarity via linear scan + min-heap (default k = 10), with glob-based key filtering; query vector goes through the embedding cache so cache HIT avoids the API
• Add VSIM <key1> <key2> — cosine similarity between two stored vectors, (nil) on any missing vector, explicit dim-mismatch error
• Add VGET <key> — returns the stored vector as a JSON array (debug helper), returns a defensive copy to prevent callers from mutating Entry.Vector

UPDATE

• Plain Set() now clears Entry.Vector on overwrite — a key re-set without the VECTOR flag drops the stale embedding since the underlying text has changed
• Extend AOF with addToAOFWithVector / writeAOF shared helpers; vector field is only emitted when len(vec) > 0 to preserve backward compatibility with existing records

Files Changed

Generated by SKILL