fix(python/redis): restore vector search broken by redisvl>=0.5 API change and KeyError on include_vectors=False

[Patrick Ribbsaeter (patrickswedish)]

Summary

Fixes #13896.

Vector search (FT.SEARCH) via the Python Redis connector was completely broken due to two independent bugs. Both have been fixed in this PR.

---

Bug 1 — process_results() API break with redisvl >= 0.5

Root cause

redisvl 0.5.0 changed the signature of process_results():

# redisvl < 0.5 (old API)
process_results(results, query, storage_type: StorageType)

# redisvl >= 0.5 (new API)
process_results(results, query, schema: IndexSchema)

The pyproject.toml pin redisvl ~= 0.4 resolves to >=0.4, <1.0 under PEP 440, so uv.lock silently picks up redisvl 0.15.x where the old API is gone. Every call to collection.search() raised:

VectorSearchExecutionException: An error occurred during the search:
'StorageType' object has no attribute 'index'

Fix

Detect the redisvl API version at runtime via inspect.signature and call process_results() with the appropriate third argument. A fallback swap is included in case the signature inspection is ambiguous:

sig = inspect.signature(process_results)
params = list(sig.parameters.values())
if "IndexSchema" in annotation_name or "Schema" in annotation_name:
    # redisvl >= 0.5: build and pass an IndexSchema
    schema = IndexSchema.from_dict({...})
    processed = process_results(results, query, schema)
else:
    # redisvl < 0.5: pass the StorageType enum
    processed = process_results(results, query, STORAGE_TYPE_MAP[self.collection_type])

---

Bug 2 — KeyError in RedisHashsetCollection._deserialize_store_models_to_dicts

Root cause

The deserializer unconditionally decoded every vector field:

case "vector":
    dtype = DATATYPE_MAP_VECTOR[field.type_ or "default"]
    rec[field.name] = buffer_to_array(rec[field.name], dtype)  # KeyError!

When include_vectors=False (the SDK default for search), Redis does not return vector fields in the result dict, so rec[field.name] raised KeyError: 'vector'.

Fix

Guard with an existence check before decoding:

case "vector":
    storage_name = field.storage_name or field.name
    if storage_name in rec:
        dtype = DATATYPE_MAP_VECTOR[field.type_ or "default"]
        rec[field.name] = buffer_to_array(rec[storage_name], dtype)
    else:
        rec[field.name] = None  # vector not requested — leave as None

---

Impact

Both bugs affected every Python user using the Redis connector for vector search:

• Bug 1 blocked all search calls unconditionally.
• Bug 2 blocked hashset search whenever include_vectors=False (the default).

The connector was written against redisvl 0.4.x; the dependency pin allows installation of redisvl 0.15.x where the API changed.

---

Testing

• Traced both code paths manually against the redisvl 0.15.0 source.
• The fix is backward-compatible: the runtime API detection correctly handles both old (<0.5) and new (>=0.5) redisvl versions without requiring a version pin change.
• Existing upsert/get/delete integration tests are unaffected.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Patrick Ribbsaeter (patrickswedish)]

Hi team! 👋

This PR fixes two independent bugs that break all Redis vector search for Python users running redisvl >= 0.5:

1. process_results() API break — redisvl 0.5.0 changed the third argument from StorageType to IndexSchema, but the connector still passes StorageType, crashing every single search call with AttributeError
2. KeyError on include_vectors=False — the hashset deserializer unconditionally accesses rec[field.name] for vector fields, but Redis omits those fields when include_vectors=False (the SDK default)

Both fixes are backward-compatible — runtime API sniffing handles both old and new redisvl versions. Happy to add tests or address any review feedback!