Answer internal knowledge questions from retrieved chunks

Goal

Build traceable retrieval-augmented Q&A over an authorized document set.

Steps

1. Chunk documents by heading as in the notebook, retaining heading, text, and source-link metadata. [rag]
2. Embed with Voyage Client.embed(...), retrieve top-k by cosine similarity, and pass only those chunks to Claude. [rag]
3. Require answers to rely only on <document> evidence and list sources; report insufficient evidence rather than guess. [rag]

How to verify

Check source coverage and answer correctness on a multi-chunk evaluation set, saving retrieval and end-to-end scores separately.

Troubleshooting

If a key chunk is missed, inspect boundaries, top-k, and embedding version; do not let the answer layer fabricate links.