Fix a knowledge assistant that answers confidently and wrong

Diagnose and fix my knowledge assistant (RAG). You are an engineer who has tuned retrieval systems in production.

What it answers questions about: <the documents, how many, what format>
How documents are chunked and embedded today: <chunk size, overlap, embedding model, vector store>
How many chunks are retrieved and how they're passed to the model: <top-k, any threshold>
The answer prompt: <paste>
Five failing examples: for each, the question, the chunks that were retrieved, the answer it gave, and what it should have said: <paste>

Do this:
1. For each failing example: was the failure in retrieval (right chunk not retrieved), in generation (chunk present but answer wrong), or in refusal (should have said it doesn't know)? Show your reasoning from the chunks.
2. Chunking: what to change and why, with a specific recommendation for size, overlap and whether to split on headings.
3. Retrieval: top-k, a similarity threshold below which the model is not called, metadata filters, and whether hybrid (keyword + vector) search would help here.
4. The answer prompt, rewritten: answer only from excerpts, cite the source per claim, exact refusal wording, language handling.
5. A regression set: 15 questions (10 answerable, 5 not) with expected outcomes, to run after every change.
6. The one change to make first, and how I'll know it worked.