RAG and internal document question answering
Private question-answering over your own documents with citations, access controls, and honest handling of the questions the source material cannot answer.
A knowledge system indexes your documents so people can ask questions in plain language and get an answer with a citation to the source. The common technique is retrieval-augmented generation: find the relevant passages, then have a model answer using only those. It is for teams losing hours to hunting through shared drives for the right SOP, contract, spec, or policy.
The problem
Company knowledge is spread across a shared drive, a wiki nobody updates, several inboxes, and two long-tenured employees. New hires spend weeks learning where things are rather than how to do the job.
- Finding the current version of a policy means asking someone who has been here a while.
- Search returns filenames, not answers, and half the filenames are ambiguous.
- Three versions of the same document exist and only one is correct.
- People answer from memory because looking it up takes too long, and memory is sometimes wrong.
- A generic assistant will answer confidently about your policies without having read them.
What we build
An approved source set
Deciding which documents are authoritative before indexing anything. Indexing a shared drive wholesale means confidently citing a superseded draft.
Retrieval that fits the documents
Chunking and search tuned to your material, usually combining keyword and semantic search, because pure vector search misses exact terms like part numbers and clause references.
Answers with citations
Every answer links to the source passage. Without a citation the user cannot verify it, and an unverifiable answer is a liability.
Permission-aware results
Retrieval filtered by the identity of the person asking, so the system never surfaces a document they could not otherwise open.
Honest failure
When the documents do not contain the answer, the system says so. A knowledge system that guesses is worse than no knowledge system.
A refresh pipeline
Re-indexing as documents change, with removals honoured, so answers do not drift from what is currently true.
What you get at handoff
- Indexing and retrieval code in your repository.
- The document set and refresh schedule documented.
- The permission model written down and tested.
- An evaluation set of real questions with known correct answers.
- Training on what the system is reliable for, and what it is not.
Stack
Models
Claude, OpenAI, Embedding models
Retrieval
Hybrid keyword and vector search, Reranking, Chunking strategies
Storage
Postgres with pgvector, Managed vector databases, Existing search infrastructure
Sources
Google Drive, SharePoint, Confluence, PDFs, Internal wikis
Access
SSO, Per-user permission filtering, Audit logging
Related services
FAQ
Can we just upload documents to ChatGPT or Claude instead?
For a small, stable set of documents and a few people, often yes, and that is the cheaper answer. A built system earns its place when the document set changes, when different people should see different things, or when you need citations and an audit trail.
How do we stop it from making things up?
Constrain the answer to retrieved passages, require a citation, and build in an explicit path for saying the answer is not in the documents. Most confident fabrication comes from a system that has no way to say it does not know.
Does it respect our existing permissions?
It has to, and that is designed in rather than added later. Retrieval is filtered by the identity of the person asking, so nothing surfaces that they could not open directly. Retrofitting permissions onto an index built without them is painful.
What about the documents that are out of date?
Deciding what is authoritative is part of the project, and it often surfaces that nobody knew which version was current. That is a useful finding. Indexing everything indiscriminately produces a system that cites old drafts with total confidence.
How do we know the answers are good?
The same way as any other AI system: an evaluation set of real questions with known correct answers, scored before rollout and re-run after changes. Citations also let users verify individual answers themselves, which surfaces problems early.
Ready to scope Knowledge Systems?
Send the workflow, tool stack, or reporting problem. We will tell you what should be automated, what should stay manual, and what is worth building first.