The problem
The answers exist — in Notion, Drive, tickets, and old threads — but finding them costs senior people’s time. Generic chatbots bolted onto search hallucinate confidently and ignore permissions.
The system
A retrieval pipeline tuned to your corpus (chunking, hybrid search, reranking), generation constrained to retrieved evidence with citations on every claim, permission filters applied at query time, and an abstain path — "not in the knowledge base" is a valid answer.
How it's built
- Connectors with incremental sync; document-level ACLs enforced in retrieval
- Hybrid retrieval + reranking tuned on your real questions
- Citation-required generation; answer-quality eval set built from day one
- Feedback loop: thumbs-down becomes an eval case
Delivery
Sprint indexes a scoped corpus and measures answer quality on your top 50 real questions; Build extends coverage.
What to expect
- Time-to-answer for internal questions drops from hours to seconds
- Every answer carries sources — trust is inspectable
- Hallucination rate measured and reported, not asserted