JevMade Sign in
← Back to guides

JevMade field notes / Worked PDF retrieval tutorial

Find possible answer pages in a long report

PageIndex shows how Jev can search a report's outline, keep several promising routes and check likely answer pages.

Original by PageIndex / VectifyAIRetrieval

Listen to this guide

JevMade’s plain-English explanation

0:00 /

Our summary

A long report may contain an answer but be too large for an AI to read in one request. PageIndex's tutorial turns the report into an outline of sections and smaller sections. Jev compares those choices to find pages that might answer a question, rather than writing an answer itself.

PageIndex prepares the outline and summaries. Jev's Choice question compares possible routes, while ordinary code keeps three promising paths instead of just the winner. After choosing candidate pages, Jev's Noul question checks each page separately: does this text state information that answers the question? The script returns page numbers and excerpts.

The example still returns two candidates when every page falls below its checking cutoff. Readers must therefore inspect the evidence rather than trust a returned page as confirmation. The longer-document version uploads the PDF to PageIndex and sends summaries and page text to TypeSafe. Two reported successes are examples, not a broad accuracy study.

Key takeaways

  1. An outline lets a search narrow its attention before reading page text, but summaries and early route choices can still miss evidence.
  2. Keep several promising routes. Checking a page separately is different from picking the best page among unsuitable options.
  3. Review returned pages and any surrounding evidence yourself. This example does not stop returning candidates when every check is weak.

Source and code review, not an independently run search. Page text is shortened before checking. The script limits batches of pages, but does not apply the same size checks to section choices; an unusually large choice could exceed the service's limits. Both hosted services require keys. The two annual-report successes are maker-reported, not a broad accuracy study. This explanation adapts VectifyAI's Apache-2.0 tutorial in our own words.

GitHub tutorial · Original published

Read the original guide Opens the author’s site in a new tab.

Keep this for later

Sign in to bookmark experiments, guides and videos, and keep notes only you can see.

Continue to sign in

We’ll bring you back to this listing.