Python: Bug: Vector search (FT.SEARCH) via the Python Redis connector is broken #13896

Describe the bug

Vector search (FT.SEARCH) via the Python Redis connector is completely broken due to two issues in the search path:

1. RedisCollection._inner_search (line 324) passes a StorageType enum to redisvl's process_results(), which since redisvl 0.5.0 expects an IndexSchema object. Every search call raises AttributeError: 'StorageType' object has no attribute 'index'.

2. RedisHashsetCollection._deserialize_store_models_to_dicts (line 620) unconditionally calls buffer_to_array(rec[field.name], dtype) for every vector field. When include_vectors=False (the default for search), the vector field is absent from the result dict, causing KeyError: 'vector'.

Issue 1 blocks all search calls. Issue 2 blocks hashset search even after issue 1 is fixed, unless the user explicitly passes include_vectors=True.

Expected behavior

collection.search(vector=[...], top=N) should execute FT.SEARCH and return results. With include_vectors=False (default), results should contain data fields but not vector fields.

To Reproduce

Prerequisites: a Redis Stack server reachable on localhost:6379, redis-cli, and a working Python SK dev environment.

1. Install Python SK per python/DEV_SETUP.md:

   cd python && make install-sk PYTHON_VERSION=3.12

2. Start Redis Stack on port 6379:

   podman run -d -p 6379:6379 docker.io/redis/redis-stack:latest

3. From python/, run the following script:

   # repro_search.py
   import asyncio
   from dataclasses import dataclass, field
   from typing import Annotated
   from uuid import uuid4
   
   from semantic_kernel.connectors.redis import RedisHashsetCollection
   from semantic_kernel.data.vector import VectorStoreField, vectorstoremodel
   
   
   @vectorstoremodel
   @dataclass
   class MyModel:
       vector: Annotated[
           list[float] | None,
           VectorStoreField("vector", index_kind="hnsw", dimensions=3,
                            distance_function="cosine_similarity", type="float"),
       ] = None
       id: Annotated[str, VectorStoreField("key", type="str")] = field(
           default_factory=lambda: str(uuid4())
       )
       content: Annotated[str, VectorStoreField("data", type="str")] = "hello"
   
   
   async def main():
       async with RedisHashsetCollection(
           record_type=MyModel,
           collection_name="repro_search",
           prefix_collection_name_to_key_names=True,
       ) as col:
           await col.ensure_collection_deleted()
           await col.ensure_collection_exists()
           await col.upsert([MyModel(id="1", content="test", vector=[0.1, 0.2, 0.3])])
   
           # This raises AttributeError: 'StorageType' object has no attribute 'index'
           results = await col.search(vector=[0.1, 0.2, 0.3], top=1)
           async for r in results.results:
               print(r.record.id, r.record.content)
   
           await col.ensure_collection_deleted()
   
   
   asyncio.run(main())

   REDIS_CONNECTION_STRING="redis://localhost:6379" uv run python repro_search.py

   Output:

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

Root cause

The connector was written against redisvl 0.4.x where process_results(results, query, storage_type: StorageType) accepted a StorageType enum. Upstream redisvl 0.5.0 changed the third parameter to schema: IndexSchema. The pyproject.toml pin redisvl ~= 0.4 resolves in PEP 440 to >=0.4, <1.0, so uv.lock picks up redisvl 0.15.0 where the API is incompatible. Tightening the pin to <0.5 is not possible because redisvl 0.4.x requires redis <6 while SK requires redis >=6.

The KeyError: 'vector' in the hashset deserializer is a separate code defect: the deserialization loop does not check whether the vector field is present in the result dict before attempting to decode it.

Platform

• Language: Python
• Source: main branch of microsoft/semantic-kernel (SK version 1.41.2)
• redis-py version: 6.4.0
• redisvl version: 0.15.0 (resolved from ~= 0.4 pin)
• Backend tested: Redis Stack 7.4.7 (RediSearch v21020, ReJSON v20809)
• IDE: Kiro
• OS: macOS 25.4.0 (Darwin), arm64

Additional context

This affects every Python user who installs SK from the committed uv.lock or resolves dependencies fresh. The redisvl API break happened at 0.5.0; the committed lockfile resolves 0.15.0. No version of redisvl that is compatible with both the SK code and the SK redis >=6 requirement exists — the code must be updated to the new redisvl >=0.5 API.

The KeyError in the hashset deserializer is independent of the redisvl version and affects any search call with include_vectors=False (the default).

Test coverage gap

Tried to use the Redis connector for vector search and it didn't work. The existing integration tests (tests/integration/memory/test_vector_store.py) only cover single-record upsert → get → delete and never call collection.search(), so these bugs have had zero test coverage. Added more tests covering the full public surface — vector search, batch CRUD, filters, paging, include_vectors, prefix mode, etc. — and that's how these issues were found. The new tests should be included alongside the fix to prevent regression.