Read long PDFs with AI
Split a local document into page-based PDF and text chunks, read them in sequence and consolidate source-linked notes, checking important visuals against the original.
Read long PDFs with AI through a sequence you can track rather than one oversized request. The pdf-chunk-reader skill splits a local document into PDF chunks and extracted text, with a manifest listing their page ranges. The agent reads each part, records findings and combines the notes into a structured summary.
Splitting does not establish that every passage was understood or extracted correctly. Scans may yield little text, and equations or tables can lose important details. Inspect consequential page images, label unreadable material and treat the synthesis as a draft to verify.
What it does
- Split an authorised report into manageable page ranges
- Track reading order through the manifest
- Keep extracted text beside the matching PDF chunks
- Record important numbers with units and source pages
- Combine checked notes into a structured summary
- Revisit figures and tables in their original context
Install and use
- Prepare the complete folder, Python3.8 or later and separately installed pypdf. Use an agent that can run commands and read local files. Provide an authorised PDF and a fresh writable output directory to avoid replacing earlier output.
- Give a specific request: “Read this hundred-page water-use report in sections. Record key claims and figures with page numbers, and flag unreadable tables.” Define the question the notes should answer.
- Read the manifest first, then process each chunk in order. Further paginate large text files when necessary. Missing extraction is a gap to investigate, not evidence that the page contains nothing.
- Consolidate the notes and verify central claims against the original pages. The manifest can retain source paths and PDF metadata; keep or share the working folder only as authorised.
FAQ
Source
Made by Liu Fang. Original skill: pdf-chunk-reader, published on GitHub under the MIT licence.
View on GitHub