Fix document retrieval in our AI chatbot (PDF Q&A)

Users can upload PDFs without problems, but our chatbot sometimes answers "I couldn't find that information" even when the answer is clearly in the document. Other times it gives a generic answer instead of using the file. Rephrasing the same question changes the result.

Example: if a document explains our refund policy, "When can I request a refund?" should find that section and answer with a source.

What we need:
- Find where retrieval fails: PDF extraction, chunking, embeddings, vector search or the context sent to the model
- Fix it so differently worded questions find the right section
- Show the correct document and page in answers
- Say clearly when the documents do not contain the answer
- Test against an agreed question list

Keep changes limited to what is needed. Deliverables: the fix, before/after results and a short root-cause note.