redisvector

Overview ¶

Package redisvector provides a Redis-backed VectorStore adapter using the RediSearch module (Redis Stack) for KNN vector similarity search.

It is built on the official go-redis client (github.com/redis/go-redis/v9). The previous hand-rolled TCP implementation had two fatal correctness bugs:

1. It read replies with a single 64KB net.Conn.Read, so any FT.SEARCH result larger than 64KB (trivial with vectors/metadata) was silently truncated and mis-parsed.
2. It encoded vectors as a comma-separated ASCII string ("0.1,0.2,..."), but RediSearch VECTOR fields and the KNN $BLOB parameter require raw little-endian float32 bytes. The old adapter could never return a correct match against a real RediSearch index.

This rewrite delegates framing to go-redis and encodes vectors as binary float32 blobs. RESP2 is forced (Protocol: 2) so FT.SEARCH replies have the stable positional array shape that parseSearchReply understands.

Testing note ¶

RediSearch (FT.*) is a Redis module and is NOT emulated by miniredis, so the unit tests here cover the pure, wire-critical logic: binary vector encoding, search-query construction, and reply parsing. End-to-end coverage requires a live Redis Stack instance; set REDISVECTOR_ADDR to run TestIntegration.