AI-Powered Enterprise Search Consulting

It is the work of putting cited question answering on top of company data using hybrid retrieval, query-time ACL, and evals — so people can ask in natural language instead of guessing keywords. Ranked links stay until evals say an answer is safe. We implement in the client cloud in four weeks standard. You own the IP. You pay inference. There is no shared training.

Not as a dump-the-drive chatbot. Production replacement is hybrid search (BM25 plus embeddings) with per-user ACL at query time, required citations, and refuse-when-empty. A chat window that ignores the search index will miss identifiers and leak files. We treat the current search engine as the permission and recall baseline to beat, not as something to switch off in week one.

Yes. An answer without a source is a defect. The UI can lead with a cited paragraph and still show ranked documents. Some queries should stay link-only — especially when retrieval is weak. We decide that per query class using the golden set, not with a global “always answer” flag.

Per-user ACL at query time, matching the current search index. If today’s search cannot return a file to that user, the generator cannot see it. We test parity with restricted accounts in week 3. Broken inheritance, guest accounts, and nested groups are discovery items, not afterthoughts.

A standard cited-Q&A layer is four weeks for a scoped corpus and identity mapping. Tenant-wide replacement of a mature search stack may need more than one four-week slice; we say so in discovery rather than promising a rip-and-replace. Financial services typically 8–12 weeks. Healthcare typically 10–14. Thirty days on-call after each handover.

They stay zero unless a new source is in scope. Refuse-when-empty is the correct behavior. Generating a plausible answer on a zero-result query is how “AI search” loses trust in week two. We include zero-result examples in the golden set so CI fails if someone removes the refusal path.

Only if it cannot support hybrid retrieval and query-time ACL in your cloud. If it can, we add the answer layer and evals on top. This is implementation consulting, not a mandatory vendor swap. Model-agnostic generators; client pays inference.

You own the IP: query classifiers, fusion config, prompts, evals, traces, runbooks, running in the client cloud. Fixed-scope implementation fee plus platform fee. No shared training. Thirty days on-call included. Further corpus expansion is a new fixed scope.