Build a Docs Q&A Agent with Source Citations and an Eval Script (Python, Claude API)

We have about 60 help-center articles in Markdown. We want a small Python agent that answers questions using only those articles. It must:cite the article and section it used for every answer, say "I don't know" when the articles don't cover the question, instead of guessing, come with an eval script that runs 30 test questions (we provide them) and reports accuracy, how often it correctly says "I don't know", and the average cost per answer. Use the Claude API. Pick the retrieval setup you think fits and tell us why. Deliverables: a GitHub repo, a README with setup steps, and the eval results. Screening questions 1.Share an AI agent or RAG system you built that real people used. What went wrong once it was live? 2.How will you stop the agent from answering when the articles don't cover the question?