Paper Copilot

A local AI agent that reads a research paper (PDF) and produces a structured, Notion-ready markdown summary. Built with LangChain and Ollama — all processing happens on your machine, no data leaves it.

What it does

Drop a PDF, get a markdown summary containing:

• Metadata — title, authors, venue, year, DOI
• Overview — what the paper is about
• Contribution — what the authors propose and why it matters
• State of the Art — related work landscape
• Methodology — high-level overview + detailed walkthrough with formulae
• Evaluation — datasets, metrics, baselines
• Key Results — main findings as bullet points
• References Analysis — pie chart of venue distribution
• Figures — relevant images extracted from the PDF with captions

Quickstart

Prerequisites

• Python 3.12+
• Ollama running locally
• A model that supports tool calling pulled via Ollama

Setup

# Install uv (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone and install
git clone git@github.com:MarcosRodrigoT/Paper-Copilot.git
cd Paper-Copilot
uv sync

# Pull an Ollama model
ollama pull gpt-oss

Run via CLI

uv run python -m src.agent input/paper.pdf

Run via web UI

uv run streamlit run app.py

Then open http://localhost:8501, drag-and-drop a PDF, and click Generate Summary.

Architecture

PDF ──► extract_text ──────────────────────────────► LLM (per-section) ──► save_markdown ──► summary.md
    ├─► extract_images ──► figure selection (LLM) ─┘                       + images/
    └─► parse_references ──► generate_chart ────────────────────────────────┘

The pipeline uses a hybrid approach:

• Code handles deterministic steps: PDF parsing, image extraction, reference parsing, chart generation, markdown assembly
• LLM handles creative steps: writing each section summary individually with focused prompts and only the relevant paper text

This is more reliable than a pure ReAct agent because local models tend to stop calling tools after seeing large results.

Configuration

Edit src/config.py to change the model or paths:

OLLAMA_MODEL = "gpt-oss"              # Any Ollama model
OLLAMA_BASE_URL = "http://localhost:11434"

Good model choices for 2x RTX 4090:

• gpt-oss — built for structured outputs and agentic use
• mistral-small — best documented for function calling
• mistral-small3.2 — improved function calling, vision capable

Project structure

├── app.py                  # Streamlit web UI
├── src/
│   ├── agent.py            # Orchestration pipeline
│   ├── config.py           # Model and path settings
│   ├── prompts.py          # Per-section LLM prompts
│   └── tools/
│       ├── extract_text.py     # PDF → text sections
│       ├── extract_images.py   # PDF → image files + captions
│       ├── parse_references.py # References → structured data
│       ├── generate_chart.py   # Venue data → pie chart PNG
│       └── save_markdown.py    # Everything → final .md
├── input/                  # Drop PDFs here
├── output/                 # Generated summaries
└── docs/                   # Step-by-step explanations

Documentation

The docs/ folder contains step-by-step explanations of every component, written as a learning resource:

Doc	Topic
00-project-overview.md	Project summary and plan
01-setup.md	uv, Ollama, and project setup
02-pdf-extraction.md	How PDF text extraction works
02b-image-extraction.md	How PDF image extraction works
03-tools.md	All five tools and how they connect
04-agent.md	The agent loop and LLM orchestration
05-ui.md	Streamlit UI integration