MISSION 02
Grounded Knowledge
Ask questions of your policies, manuals, contracts and tickets. Answers link to the paragraph they came from, and use only files the person asking is allowed to open.
The problem
A plain chatbot answers confidently whether or not it knows. In a policy, claims or support setting, a wrong answer stated well is worse than no answer.
Most document search also ignores permissions. If the index can read the HR folder, so can everyone who asks it a question.
What we build
- Connectors for where your documents already live: shared drives, wikis, ticket systems, PDFs and scans.
- Ingestion that keeps structure, versions and access rules with every passage.
- Retrieval with pgvector and open embedding models, filtered by the asker’s permissions before the model sees anything.
- Answers with citations to the source paragraph, and a plain “not in your documents” when that’s the truth.
- Tests built from real questions your team asks, including ones the documents can’t answer.
Reference architecture
What you own at the end
- Ingestion and retrieval code, in your repository
- A question set with expected answers and sources
- Runbooks for re-indexing and adding new sources
- Dashboards for answer quality and unanswered questions
When it’s not the right fit
If you have a few hundred documents that rarely change, good search may be all you need. We’ll set that up instead.
Talk about Grounded Knowledge
Tell us roughly what you’re trying to build. A few sentences is plenty. We’ll reply by email.
Start a conversation