Chat With Your Documents (demo project)
Demo on sample data. Not a client project. Documents are a fictional store's returns, shipping and warranty policy. Answers shown come from the retrieval code and, in the generation path, a local test stand-in, not a live AI model.
The problem
Staff and customers ask the same questions that are already answered in policy documents nobody can search.
What it does
The bot splits your documents into passages, finds the ones that match a question, and answers from them with the document and line cited. If nothing matches, it says it couldn't find the answer and suggests contacting support.
Design choices
- Answers come only from retrieved passages, and the AI is instructed to say "I don't know" otherwise.
- Citations let a reader check every answer.
- Plain keyword retrieval in the demo; a paid build uses embeddings and a vector store for larger document sets.
What we tested
Five automated checks pass (returns question finds the returns policy, shipping cost finds shipping, warranty question finds warranty, refund timing finds the 5-day line, nonsense gets the not-found message). The generation path ran through the real Anthropic Python library against a local stand-in, proving the wiring.
What we did not test
Answer quality with a live AI model, large document sets, or scanned PDFs. A client job includes a test sheet of the client's real questions with honest results.
Files: `demos/2-docs-qa-bot/`.