Repository navigation
bug(vector_stores/redis): search score hardcoded for cosine, breaks L2 ranking #7249
Description
Description
Same class of bug as #6546 / #6547 / #7229, but in the Redis vector store.
RedisDB.search hardcodes:
score=max(0.0, 1.0 - float(result["vector_distance"]))That mapping is only correct for cosine. create_col(..., distance=...) already accepts a metric and writes it into the RedisVL schema (distance_metric), but search scoring never reads that metric — L2 (and similar unbounded distances) collapse almost every hit with d > 1 to 0.0, breaking ranking and score thresholds.
Also, RedisDB.__init__ / RedisDBConfig have no distance field, so the default index is always created as cosine in DEFAULT_FIELDS, and there is nowhere for operators to declare the metric used for scoring.
Steps to Reproduce
- Create a Redis collection with a non-cosine metric:
store.create_col(name="demo", vector_size=768, distance="l2")
- Insert vectors and search.
- Observe returned scores: neighbors with L2 distance
> 1become0.0under1 - distance.
Expected Behavior
- Config / constructor can declare a distance metric (default cosine).
create_col(..., distance=...)persists the metric for scoring, not only forFT.CREATE.- Search converts distance → similarity in a metric-aware way (same approach as fix(vector_stores/s3_vectors): make search score metric-aware #6547):
- cosine / ip:
max(0, 1 - d) - l2:
1 / (1 + d)so scores stay in(0, 1]and preserve ordering
- cosine / ip:
Additional Context
- Sibling of S3 Vectors search score is hardcoded for cosine, breaks euclidean ranking #6546 (S3), bug(vector_stores/valkey): search score hardcoded for cosine, breaks L2 ranking #7229 (Valkey), bug(vector_stores/vertex_ai): search score hardcoded for cosine, breaks L2-style indexes #7232 (Vertex AI). Prefer sequential small PRs.
- Happy to open a PR mirroring fix(vector_stores/s3_vectors): make search score metric-aware #6547 / fix(vector_stores/valkey): make search score metric-aware #7230 once this has the
acceptedlabel.
Activity
fix(vector_stores/redis): make search score metric-aware
ktz03 commentedon Sep 8, 2026
PR ready for this issue (Redis sibling of #6546/#7229). It will auto-reopen once labeled
accepted.@kartik-mem0 when convenient, could you add
accepted? Preferred sequence: Valkey #7229 → this Redis fix → Weaviate #7250 / Vertex #7232.fix(vector_stores/redis): make search score metric-aware
ktz03 commentedon Sep 8, 2026
PR ready for this issue (Redis sibling of #6546/#7229). It will auto-reopen once labeled
accepted.@kartik-mem0 when convenient, could you add
accepted? Preferred sequence: Valkey #7229 → this Redis fix → Weaviate #7250 / Vertex #7232.blazingretinol commentedon Sep 25, 2026
Traced this on main and could not confirm it as a live bug. RedisDBConfig has no distance field, extra keys are rejected, and RedisDB.init always builds the index with distance_metric cosine in DEFAULT_FIELDS. Memory never calls create_col with a distance argument, so mem0's Redis store is cosine only today: the L2 score collapse you describe can't happen through any setting we expose. create_col(distance=...) is real code but only reachable by calling it directly on the store object, bypassing config entirely. PRs #7251 and #7253 are identical and would add the missing config plus the metric-aware scoring together, that combination is what this actually needs.