RAG Consulting

RAG consulting is paid help to design, build, and hand over a retrieval-augmented generation system that answers from your documents instead of the model’s memory. The engagement covers corpus design, hybrid retrieval (BM25 plus embeddings), per-user ACL at query time, required citations, refuse-when-empty, and golden sets plus CI evals. Work runs in the client cloud. You own the IP. A standard implementation is four weeks, then 30 days on-call.

You need RAG when the source of truth is documents that change and must stay permissioned. You need fine-tuning when the model must acquire a skill, format, or tone that prompting cannot hold. Most enterprise Q&A, policy, and search problems are RAG problems. Some systems use both: RAG for facts, a light fine-tune for voice. We decide in discovery; we do not fine-tune as a default.

A standard RAG implementation is four weeks: discovery week 1, environments week 2, shadow-mode pilot week 3, handover week 4. Financial services typically takes 8–12 weeks because of model-risk review. Healthcare typically takes 10–14 weeks. Ecommerce typically takes 6–10 weeks. Scope is fixed in writing. Thirty days on-call is included after handover.

No. There is no shared training. Documents are indexed in the client cloud for retrieval; they are not used to train a foundation model we reuse for anyone else. You own weights, datasets, eval suites, and runbooks. We are model-agnostic. You pay inference to the provider. That split is in the statement of work.

Hybrid retrieval so the right passage is in context, a refuse-when-empty rule so the model cannot invent when the index misses, and a citation requirement so every claim points at a span. Then golden sets plus CI evals measure faithfulness on real questions. A larger model without those controls still fabricates. Hallucination prevention is a system, not a prompt trick.

Yes, if ACL is enforced at query time. Document ACLs are stored with the chunks; the retriever filters to the asking user’s live entitlements before generation. A single service account over the corpus will leak restricted files. We wire identity you already run — Entra ID, Okta, or equivalent — rather than a parallel permission store.

You own the IP: indexes, chunking code, prompts, eval suites, traces, and runbooks, running in the client cloud. There is no proprietary runtime you rent forever. Commercial terms are a fixed-scope implementation fee plus a platform fee, quoted in writing. You pay inference. Thirty days on-call is included so your team can operate the system without us in the loop.

It is a fixed-scope implementation fee plus a platform fee, quoted in writing after discovery of corpus size, ACL complexity, and eval burden. We do not mark up tokens; the client pays inference. Model choice is yours. Standard scope is four weeks to handover. If the corpus or control requirements exceed that, we re-scope before build rather than absorbing it into a vague retainer.