Selective context window

[xiyivan]

Is it possible to select context attached based on chapter/sections? It will be very helpful in controling usage when reading textbooks or long reports

[dralkh]

Greetings xiyivan, thanks for the suggestion. I agree this could make the context much more focused and enhance response greatly.

The main challenge is that PDFs vary a lot in structure, metadata, parsing quality, and heading detection, so chapter or section selection may not always be reliable. That said, I’ll look into adding checks and possibly semantic matching in detecting these headings effectively to the added context.

What type of headings you have in mind to triangulate? and If you have any files I can test against, it would be helpful to match your use case.

[xiyivan]

Hi dralkh, thanks for the reply. I completely agree that PDFs tend to be pretty inconsistent in organizing their bookmarks.

Most of books I downloaded with chapter-based bookmarks use a tree like structure:

• Chapter 1 Title
  • 1.1 Section Title
• Chapter 2 Title
  • 2.1 Section Title

Sometimes the text Chapter is omitted and the bookmarks looks like below:

• 1 Title
  • 1.1 Section Title
• 2 Title
  • 2.1 Section Title

I would say even a relative simple parser that uses the existing bookmark tree structure and basic regex matching would be a huge step forward and cover a lot of ground.

With that said, for pdfs without proper bookmarks. One method I can think of is to perform OCR test on the content section of the PDF. I found several repos like huridocs/pdf-table-of-contents-extractor and kszenes/tocPDF deal with similar issues.

Hope this helps. Let me know what you think.

[dralkh]

Thank you for your response and examples provided, will look to into and test against repos supplied and other methods. I propose modal in context field for flexibility to choose array of certain section or similar match to applied keywords with some set of known ones for typical papers. and furthermore any of context item that passed selection to be founded to be highlighted in context menu. and others that didn't may pass the full paper to the context and may be disabled if not from within configuration.

[xiyivan]

That sounds great.

Also just a quick thought: even just allowing selection based on page number range would be a very helpful function. (as the bookmarks just help in locating the page number)

Looking forward to see this develop.