Enterprise AI Knowledge Management Consulting

It is RAG over internal knowledge — wikis, SharePoint, approved drives — so employees ask questions and get cited answers from current pages, with per-user ACL at query time. It is not a new knowledge-base product. We implement hybrid retrieval, refuse-when-empty, golden sets plus CI evals, in the client cloud, in four weeks standard. You own the IP. There is no shared training.

Yes, if connectors preserve ACL and the first corpus is curated. We index approved spaces, apply per-user ACL at query time, use hybrid retrieval (BM25 plus embeddings), and require citations back to the page. Indexing every site collection on day one retrieves drafts and orphaned copies. Discovery names the first two or three sources that actually answer the questions.

Citations make staleness visible, owners stay responsible for pages, and CI evals fail when a golden question’s gold passage moved or died. We do not silently “correct” a page from model memory. Refuse-when-empty still applies if retrieval cannot find a current span. Knowledge management remains a publishing problem; RAG does not replace editors.

It must not. Per-user ACL at query time filters chunks to the asking user’s live entitlements before generation. A crawler service account that indexes everything and then “hides” files in the UI will leak. We test this in week 3 with accounts that should miss restricted libraries.

Four weeks standard: discovery, environments, shadow-mode, handover, then 30 days on-call. Financial services typically 8–12 weeks. Healthcare typically 10–14. Ecommerce typically 6–10. If identity or connector work will not fit, we re-scope in writing. Client owns IP in the client cloud.

Yes. Generative AI reads the knowledge base; it does not replace it. Pages still need owners, review dates, and a place of record. RAG that cites a heading is useful because people can open the source. If the wiki is abandoned, fix publishing first or the system will cite abandoned pages fluently.

We are model-agnostic: Anthropic, OpenAI, Google, Mistral, open-source, or a client fine-tune. Work runs in your cloud. You pay inference. No shared training. Model choice is scored on your golden set, not on a brand preference. Swaps after handover go through the same CI evals.

Fixed-scope implementation fee plus platform fee, quoted after we know source count, ACL complexity, and eval size. You pay inference. You own indexes, connectors, and evals. Thirty days on-call is included. We do not sell a per-seat knowledge chatbot license as the implementation.