SemanticCache vector index algorithm is hardcoded to FLAT; no way to configure HNSW

[vishal-bala]

Current behavior

SemanticCacheIndexSchema.from_params (redisvl/extensions/cache/llm/schema.py:115) hardcodes the vector field's algorithm to "flat":

{
    "name": CACHE_VECTOR_FIELD_NAME,
    "type": "vector",
    "attrs": {
        "dims": vector_dims,
        "datatype": dtype,
        "distance_metric": "cosine",
        "algorithm": "flat",
    },
},

There is no constructor argument on SemanticCache to override this. FLAT is exact KNN — fine to the low millions of entries, latency-degrading beyond that. For large multi-tenant deployments or per-tenant player-facing caches (10M–100M entries), HNSW is required. Today the only workaround is to bypass SemanticCache entirely and build the index manually.

Proposed API

SemanticCache(
    name="cache",
    redis_url=...,
    vector_index_config={
        "algorithm": "hnsw",          # or "flat" (default)
        "m": 16,
        "ef_construction": 200,
        "ef_runtime": 10,
    },
)

A simpler form taking just algorithm: Literal["flat", "hnsw"] = "flat" plus a default HNSW parameter preset would also be acceptable for a first cut.

Compatibility notes

The algorithm is fixed at index creation; existing FLAT caches cannot be hot-swapped to HNSW. Customers wanting to switch must build a new cache (new name), re-warm, and cut over — same migration shape as a schema change. This is acceptable as long as the option exists at construction time.

Notes

Surfaced while writing a scoped semantic caching architecture spec for customers running large-scale, multi-tenant deployments.