Python: RedisCollection.ensure_collection_deleted can return successfully without deleting the index #14473

Describe the bug

RedisCollection.ensure_collection_deleted can return successfully without deleting anything.

# python/semantic_kernel/connectors/redis.py
async def collection_exists(self, **kwargs) -> bool:
    try:
        await self.redis_database.ft(self.collection_name).info()
        return True
    except Exception:
        return False

async def ensure_collection_deleted(self, **kwargs) -> None:
    exists = await self.collection_exists()
    if exists:
        await self.redis_database.ft(self.collection_name).dropindex(**kwargs)
    else:
        logger.debug("Collection does not exist, skipping deletion.")

collection_exists() reaches Redis to answer, so it can fail for reasons unrelated to whether the index is there - a timeout, an authentication failure, an unreachable node. Every one of those returns False, and ensure_collection_deleted then logs at debug level and returns normally.

The caller is told the collection is deleted. It is not. Nothing raises and nothing is logged above debug.

The same library already does this correctly

ChromaCollection.ensure_collection_deleted in python/semantic_kernel/connectors/chroma.py distinguishes the two cases:

try:
    self.client.delete_collection(name=self.collection_name)
except ValueError:
    logger.info(f"Collection {self.collection_name} could not be deleted because it doesn't exist.")
except Exception as e:
    raise VectorStoreOperationException(
        f"Failed to delete collection {self.collection_name} with error: {e}"
    ) from e

Absent is handled quietly. Anything else raises. That is the behaviour I would expect from ensure_collection_deleted on any connector, and the Redis one is the odd one out.

ChromaCollection.collection_exists has the same except Exception: return False shape, but its delete path does not depend on it, so the consequence there is smaller.

To reproduce

Call ensure_collection_deleted() on a RedisCollection while Redis is reachable at connect time but the FT.INFO call fails - a timeout is the easiest to induce. The method returns without raising and the index is still present.

Expected behavior

A failed check is not an absent collection. ensure_collection_deleted should either delete the index or raise.

Suggested fix

Mirror the Chroma connector: attempt dropindex and handle only the specific "index does not exist" error quietly, raising VectorStoreOperationException for anything else.

I am filing this rather than opening a pull request because the right fix depends on exactly which exception FT.DROPINDEX raises for a missing index in the redis-py version this pins, and I did not want to assert that without a Redis instance to confirm it against. If a maintainer can confirm the exception type, I am happy to open the PR.

Platform

Python, main as of 23 September 2026.