Native hybrid_search for the 6 remaining hybrid-capable backends

[thorwhalen]

Native hybrid_search for the 6 remaining hybrid-capable backends

PR #16 landed SupportsHybrid + the universal client-side BM25 + RRF fallback, and wired the 3 backends whose adapters already provision text-side indexing (weaviate, elasticsearch, redis) as SupportsHybrid.

The 6 other backends data/providers.yaml marks hybrid_search: true still go through the client-side fallback — correct but O(N) over the collection. Each needs its own infra wiring before its native path can be turned on. They're listed below in roughly easiest-first order.

lancedb

• Add table.create_fts_index("text") to create_collection (or lazy-create on first hybrid call).
• Define _lexical_query calling table.search(text, query_type="fts"), or use the unified table.search(text, query_type="hybrid") and skip RRF orchestration on this path.

mongodb (Atlas)

• Auto-create a $search (Atlas Search) index alongside the existing $vectorSearch index in create_collection — analogous to how _search_index_create already provisions the vector index.
• Define _lexical_query running a $search aggregation pipeline.
• Or use the full $rankFusion aggregation stage to do dense + lexical + RRF in one round-trip.

qdrant

• Add a sparse-vector config to create_collection and a sparse-vector encoder (e.g. SPLADE, BM25 sparse) at write time.
• Define _lexical_query issuing a sparse-vector query_points call, OR use query_points(prefetch=[dense, sparse], query=FusionQuery(fusion=Fusion.RRF)) for native fusion.

milvus (≥ 2.5)

• Add a BM25 function field to the collection schema in create_collection.
• Define _lexical_query issuing a client.search call against the BM25 field.
• Or use client.hybrid_search([AnnSearchRequest(...), AnnSearchRequest(...)], ranker=RRFRanker()).

pinecone

• Generate sparse vectors at write time (BM25 / SPLADE / pinecone-text).
• Either upsert with both values and sparse_values, then query with both — or use Pinecone's hosted hybrid endpoint.

turbopuffer

• Define a BM25 attribute in the namespace schema and populate it at write time.
• Use query(rank_by=("text", "BM25", text)) for the lexical side, fused via rank_by with the dense rank.

Refs

Follow-up to #16, refs #11. Each backend is its own clean piece of work — feel free to split into separate PRs.