RAG debugging worksheet: inspect retrieval and answers separately

Let us use a shared worksheet when discussing a RAG answer that misses the expected source material. The goal is to show what each stage produced before changing several settings at once.

• Question: A shareable example of the user input.
• Expected evidence: The document passage you expected the system to use.
• Retrieved material: The chunks returned, in their original order.
• Configuration: Embedding model, chunking, filters, retrieval count, and any reranking.
• Answer input: The instructions and context actually passed to generation.
• Final answer: The relevant excerpt and the specific error.

Use public documents or a small synthetic example when your corpus is private. Include document versions and a date for the test.

Which stage would you inspect next, and what observation would support changing it? Share results with the original example so the discussion remains reproducible.