AI Document Q&A Generative AI Consulting

Index approved PDFs and policies with layout-aware chunking, retrieve with BM25 plus embeddings, filter by per-user ACL at query time, generate only from those spans with page citations, and refuse when retrieval is empty. Golden sets plus CI evals gate clause questions. We implement that in the client cloud in four weeks standard. You own the IP. There is no shared training on your documents.

The user experience can be a chat window. The system is not a consumer ChatPDF wrapper. Production document Q&A has query-time ACL, hybrid retrieval, citation to page, refuse-when-empty, and evals on real clauses. Demos that upload a file to a vendor tenant and train nothing still fail those tests. We build in your cloud so the files never become someone else’s corpus.

Yes, as assistive Q&A with citations — not as an automated legal determination. Reviewers still own the conclusion. Tables, defined terms, and cross-references need layout-aware chunking and hybrid retrieval or the model will quote the wrong section. Missing clauses refuse. We will not market this as a substitute for counsel.

Refuse-when-empty, required citations to retrieved spans, and faithfulness evals on a golden set of clauses. Completing a gap from the model’s training data is the failure mode. Hybrid retrieval (BM25 plus embeddings) is how clause numbers and defined terms actually get into context. CI fails if those controls regress.

Not if per-user ACL at query time is enforced. Document library permissions travel with the chunks. A shared upload folder with no ACL is out of scope until identity is mapped. Week 3 tests restricted accounts against deal rooms and policy libraries they should miss.

Four weeks standard for a scoped document class and identity mapping, then 30 days on-call. Financial services typically 8–12 weeks. Healthcare typically 10–14. Ecommerce typically 6–10. Large backfile OCR or broken ACL is extra scope and we write it down in week 1 rather than hiding it in a retainer.

In the client cloud. You own the IP: indexes, chunking, prompts, evals, runbooks. No shared training. Model-agnostic. You pay inference. Fixed-scope implementation fee plus platform fee. We do not keep a copy of your contract library on a shared cluster.

Tables need layout-aware chunking or hybrid retrieval will miss cells. Scans need OCR in your perimeter before indexing; we will not pretend an image-only PDF is text. Discovery names which classes are digital vs scanned. If OCR is most of the work, the four-week Q&A clock starts after that pipeline exists, or we include it as written scope.