Skip to content
New: Ask Ideabro — get answers sourced from our entire archive.Try it
IDEABRO
Help Desk Solved

Long PDF + model = it ignores the middle chapters. Workarounds?

Started 5d ago 4 replies 1.3K views

Nandini Iyer

Nandini IyerOP

I'm reading a 180-page thesis from another lab, closely related to mine. I uploaded the whole PDF to an assistant with a long context window and asked it to compare their method in chapter 4 with ours.

It answered confidently, but the details came from the introduction and conclusion, not chapter 4. When I asked about a specific table in chapter 4, it described a table that doesn't exist.

The context window is supposedly big enough for the whole thing. Is the model just not reading the middle? What's the workflow people use for long documents that actually works?

4 replies

  1. Accepted answer

    Mei-Lin Zhao

    Mei-Lin Zhao

    Fitting in the window and being read carefully are different things. Long-context models recall material at the start and end of a document far more reliably than material in the middle. Two-step workflow:

    1. Extract first. "From chapter 4 only, quote every sentence that describes the method, with page numbers." No analysis yet.
    2. Answer from the quotes. Paste the quotes into a fresh request and ask your comparison question there.

    The quotes become a short, dense document the model can't skim, and you can verify them against the PDF in a minute.

  2. Yusuf Demir

    Yusuf Demir

    Also: split the PDF by chapter and upload only what you need. I do this with game design documents. A 20-page chunk with a focused question beats 180 pages with a vague one every time.

  3. Lena Hartmann

    Lena Hartmann

    The invented table is the worrying part. That's the model filling a gap with something plausible. Asking for quotes with page numbers makes that much harder to do silently. We wrote a whole piece on this: The Context Window Is Not a Memory.

  4. Nandini Iyer

    Nandini IyerOP

    Extract-then-answer worked. The quotes from chapter 4 revealed they use a different normalisation step from us, which the first answer completely missed. Lesson learned: a big desk is not a careful reader.

Join the conversation

Sign in to reply, upvote and follow this thread.

Sign inJoin free