Skip to content

Redis Stack / RediSearch

hms-import redis --url redis://host:6379 --index idx or --prefix doc: --text-field content --vector-field embedding

  • Reads documents by SCAN over the key prefix (hashes via HGETALL, JSON via JSON.GET key $), pipelined per page. --index runs FT.INFO to discover the prefix, key type and vector attribute; without RediSearch use --prefix.
  • Hash vector fields are raw little-endian float32 bytes (kept as bytes, never UTF-8 decoded); JSON vectors are arrays. Other binary hash values are stored base64 in metadata. Hash values are strings in Redis, so numeric metadata arrives as strings.
  • Password from the URL or $REDIS_PASSWORD; never logged. Unreadable JSON documents are rejected with a reason, not dropped.
  • SCAN has no stable order across mutation: resume re-reads and hash-verifies the committed prefix and restarts from the first record if it changed. Keys can be returned twice by SCAN; writes are idempotent by id.

Verification: hash/SCAN path REAL (redis-server 8, no modules): round trip, resume, missing-RediSearch error. FT.INFO parsing and JSON documents: MOCK (canned RESP2 reply and a fake client). RediSearch/RedisJSON modules themselves unverified.