Local model chore: summarize a long PDF without it leaving your laptop

The Stack

Tested Against

Prerequisites

• A laptop with ~8GB RAM
• Ollama or LM Studio for the summary step

Steps

1. 1

   Extract the document text that feeds the model (deterministic)

   When you 'attach' a PDF, the app first extracts its text to hand to the model. CI builds a small fixture PDF with reportlab, then verifies pypdf recovers the exact content, the real, deterministic ingestion step. No model, no key.

   python3 -m venv .venv
   .venv/bin/pip install -q reportlab pypdf
   .venv/bin/python - <<'EOF'
   from reportlab.pdfgen import canvas
   from pypdf import PdfReader
   
   c = canvas.Canvas("doc.pdf")
   y = 800
   for line in [
       "QUARTERLY REPORT",
       "Revenue grew 18 percent to 2.4 million dollars.",
       "Action item: renew the SOC2 audit before September.",
   ]:
       c.drawString(72, y, line); y -= 24
   c.save()
   
   text = "".join((p.extract_text() or "") for p in PdfReader("doc.pdf").pages)
   assert "Revenue grew 18 percent" in text, f"extraction missed the revenue line: {text!r}"
   assert "renew the SOC2 audit" in text, f"extraction missed the action item: {text!r}"
   print("chore2 OK: pypdf recovered the document text that feeds the local model (revenue + action item)")
   EOF

2. 2

   Summarize the extracted text on a local model (the model step, not checked by CI)

   Pipe the extracted text into Ollama or attach the PDF in LM Studio / Open WebUI and ask for 'five plain bullets and anything I need to act on.' The summary runs the model and is non-deterministic, so CI never claims it.