Enterprise RAG Implementation Consulting

You implement it as a permissioned retrieval system, not a chatbot wrapper. Hybrid retrieval (BM25 plus embeddings) over a curated corpus, per-user ACL at query time, citations required, refuse-when-empty, golden sets plus CI evals, and traces that reconstruct the prompt. We do that in the client cloud on a four-week standard clock, then hand over IP and stay on-call for 30 days.

Usually because retrieval was vector-only, a service account indexed everything, answers were allowed on empty retrieval, and nobody had a golden set in CI. Those four show up in almost every rescue. The fix is hybrid retrieval, query-time ACL, a refusal path, and eval gates — not a model swap. We audit the prototype in week 1 and write the gap list before touching production traffic.

Standard implementation is four weeks: discovery, environments, shadow-mode, handover. Financial services is typically 8–12 weeks. Healthcare is typically 10–14 weeks. Ecommerce is typically 6–10 weeks. Thirty days on-call follows handover. If ACL or corpus work will not fit four weeks, we re-scope in writing before build.

In the client cloud or VPC. Secrets, traces, and document bytes do not leave your perimeter for a shared platform index. There is no shared training. You own the IP. We are model-agnostic and you pay inference to the provider. That is the default, not an upgrade.

Per-user ACL at query time. Chunks carry the ACL of the source document; the retriever filters to the asking user’s live groups before any text is sent to the model. Hiding a file in the UI after retrieval is not a control. Group membership changes apply on the next query without a full re-embed.

A golden set of real questions with verified answers, retrieval hit@k, answer faithfulness against retrieved context, and a citation-coverage check. Those run in CI. A drop versus the last pin blocks release. We do not treat a vendor leaderboard as a substitute for your questions.

Yes, if it can store ACL metadata, run hybrid retrieval or sit beside BM25, and stay in your cloud. We are model-agnostic on embeddings and generators. If the current store cannot filter at query time, we say so in discovery rather than wrapping a leak. Client pays inference either way.

Your team operates the index, evals, and prompts. Client owns IP. Thirty days on-call is included for incidents and eval triage. Commercial terms are a fixed-scope implementation fee plus a platform fee. After the on-call window, further changes are a new fixed scope, not an open retainer unless you ask for one in writing.