Fix fallback logic in page chunk processing to prevent irrelevant results

[UdayRajSinghh]

Thank you for building such a well-structured PDF chatbot system! The codebase is clean and the three-step processing pipeline (document selection → page detection → answer generation) is really thoughtfully designed. I hope this small improvement helps make the search results even more reliable.

---

Problem

When JSON parsing fails in _process_page_chunk(), the current fallback returns the first page of the chunk regardless of relevance. This can lead to misleading search results where users receive confident answers from completely unrelated content.

Solution

• Remove the fallback that returns first_page when parsing fails
• Return empty array [] instead
• This ensures only genuinely relevant pages are included in results

Impact

• Prevents false positive results in document search
• Maintains result accuracy when some chunks fail to process
• No breaking changes to existing functionality

Fixes the edge case where parsing errors could pollute search results with irrelevant content.

[vercel]

@UdayRajSinghh is attempting to deploy a commit to the Roe AI Team on Vercel.

A member of the Team first needs to authorize it.