← All case studies
Studio internal product
AI assistant over documentation
80% of questions handled without a human
- Timeline
- 4 weeks
- Stack
- PythonFastAPIAnthropic APIQdrantNext.js
Problem
Prove on the studio's own documents the approach we offer clients: a RAG assistant answering questions over an internal knowledge base — policies, contracts, technical docs — with source citations and zero invented facts.
Solution
- An indexing pipeline: documents are split into semantic chunks and stored in a Qdrant vector database
- Answers strictly from retrieved fragments, with quotes and links to the source document
- When confidence is low, the assistant honestly says "I don't know" and calls a human instead of hallucinating
- Interfaces: a web chat and a bot in the work Telegram, one shared dialogue history
Results
- 80% of knowledge-base questions are closed without a human
- 94% accuracy on the test set (measured before launch, not eyeballed)
- The approach reproduces on any company's documents within a 2–3 week pilot