What it does

Recursive Summarizer turns any PDF into a tight recap.
It extracts every page, slices the text into 3-page chunks, asks GPT-3.5-turbo to summarize each slice, stitches the mini-summaries together, and keeps compressing until the final text fits your target length.

Quick facts

• Python script, one dependency stack (pdfplumber, openai)
• Works on scans or born-digital PDFs (OCR quality affects output)
• Defaults: ~3 200-character chunks, 10-page final cap
• Saves both an intermediate dump and a final summary.txt

How it works

1. Extract
   pdf_to_text walks every page with pdfplumber and returns raw text.

2. Chunk
   split_text splits the corpus into ~3-page slices so the model stays within context.

3. Summarize
   get_summary calls ChatCompletion once per slice, writes an intermediate file, and returns the snippet.

4. Re-curse
   If the stitched output is still too long, the script passes it back to GPT for a tighter pass until it fits.

5. Persist
   The final summary lands in summary.txt.

# one-liner demo
python recursive_summarizer.py --pdf ~/Docs/whitepaper.pdf

Tune it

| Variable | What it controls | | ----------------- | -------------------------------------------------------------- | | max_page_length | Rough characters per PDF page | | chunk_length | Characters sent to GPT in one go (max_page_length * 3 by default) | | target_length | Max size of the final summary |

Adjust these three numbers to balance cost, speed, and detail.

Caveats

• Output quality hinges on the PDF's text layer; bad scans in → garbled text out.
• Hard limits can trim nuance if you aim for ultra-short summaries.
• Each GPT call costs tokens; monster PDFs may run up your bill.

Get it

Source and MIT license live here: https://github.com/patrickmaub/recursive-summarizer

Clone, add your OPENAI_API_KEY, point at a PDF, and watch it shrink.