Hybrid Search RAG Consulting

Hybrid search RAG retrieves with BM25 and embeddings over the same chunks, fuses the candidates, re-ranks, applies per-user ACL at query time, then generates with required citations. Vector-only RAG misses exact identifiers. Keyword-only RAG misses paraphrases. Hybrid is the production default. We implement it in the client cloud in four weeks, with golden sets plus CI evals.

Better embeddings still blur exact tokens. Policy IDs, SKUs, and error strings are lexical problems. BM25 is cheap, explainable, and already good at them. We keep embeddings for semantic matches and fuse. Changing the embedding model is a later experiment, gated by the same golden set. You pay inference; we do not swap models to look busy.

Vector for paraphrase and conceptual questions. Keyword (BM25) for identifiers and rare tokens. Hybrid for production, because real query logs mix both. We prove it on your golden set rather than arguing architecture. If one side never contributes hits, we can drop it later — CI will show that, not a vendor slide.

Yes, if ACL is applied to fused candidates at query time. Running BM25 without the same filter as the vector index is how restricted filenames leak. Chunks share ACL metadata across both indexes. Empty filtered fusion refuses. We wire live identity; we do not copy groups into a nightly snapshot if we can avoid it.

Four weeks standard: discovery, dual-run, shadow-mode, handover. Financial services typically 8–12 weeks. Healthcare typically 10–14. Ecommerce typically 6–10. Thirty days on-call after handover for query and index tuning. Scope is fixed in writing. You own the IP in the client cloud.

Only if the current engine cannot do hybrid retrieval and query-time filters in your cloud. If it can, we configure it. We are model-agnostic on embeddings. Client pays inference for embedding and generation APIs. No shared training. This is implementation, not a search SaaS migration unless discovery says the engine cannot meet the contract.

Golden sets plus CI evals, stratified: identifier questions, paraphrases, and long narrative questions. We track hit@k per stratum and the contribution of BM25 vs embeddings. Faithfulness and citations are scored after generation. A change that helps paraphrase but tanks identifiers fails the gate.

A fixed-scope implementation fee plus a platform fee, quoted in writing. You pay inference. Work runs in your cloud. You own indexes, fusion code, evals, and runbooks. No token markup. Thirty days on-call is included. After that, tuning is a new fixed scope unless you extend in writing.