Case study
AI document assistant: answers from the company's own knowledge, with citations
Client under NDA
Product screenshots coming soon — the numbers are already live.
Challenge
The client's staff spent hours each week asking senior colleagues questions whose answers already existed in policy documents, contracts, and past decisions, because nobody could find anything across hundreds of files. A generic chatbot was tried and abandoned: it answered confidently and wrongly.
Solution
We built a retrieval-grounded assistant: staff ask in plain language, the system retrieves the relevant document passages, and the model answers only from those passages, with citations a click away. Questions outside the corpus get an explicit 'not in the knowledge base' response and a handoff path to a human, because a wrong answer is worse than no answer.
Architecture
We show real architecture. Agencies never do.
Documents are chunked and embedded into PostgreSQL with pgvector; a NestJS service handles retrieval, prompt assembly, and calls to the Claude API, logging every question, retrieved context, answer, and cost. An n8n pipeline keeps the index fresh as documents change. Releases must pass an evaluation suite of real staff questions before deploy, so quality is measured, not vibes.
Results
- Staff get cited answers in seconds instead of interrupting senior colleagues
- Wrong-answer risk is engineered down: grounding, refusal paths, and a pre-release evaluation gate
- Every response is logged with its cost, so the finance conversation is a dashboard, not a guess
- The client can audit exactly what the assistant told whom, and when
Want something like this?
Tell us what you are building. The founders who shipped AI document assistant: answers from the company's own knowledge, with citations will scope it and reply with a fixed price.