thom

Summarize research papers from arXiv using LLMs.

Named after René Thom (1923-2002), the French mathematician who founded catastrophe theory and won the Fields Medal in 1958.

Installation

pip install thom

Or install from source:

git clone https://github.com/thom-project/thom.git
cd thom
pip install -e .

Quick Start

Python Library

import thom

# Fetch a paper from arXiv
paper = thom.fetch_paper("2301.00001")
print(paper.title)
print(paper.abstract)

# Summarize the paper
summary = thom.summarize(paper)
print(summary.summary)
print(summary.key_points)

# Search for papers
papers = thom.search("transformer attention mechanism", max_results=5)
for p in papers:
    print(f"{p.arxiv_id}: {p.title}")

# Compare multiple papers
papers = [thom.fetch_paper(id) for id in ["2301.00001", "2301.00002"]]
analysis = thom.compare(papers)
print(analysis)

Command Line

# Summarize a paper
thom summarize 2301.00001

# Or use a URL
thom summarize https://arxiv.org/abs/2301.00001

# Fetch paper metadata only
thom fetch 2301.00001

# Search for papers
thom search "machine learning"

# Compare papers
thom compare 2301.00001 2301.00002

# Use different models
thom summarize 2301.00001 --model gpt-4o
thom summarize 2301.00001 --model claude-3-5-sonnet-20241022
thom summarize 2301.00001 --model ollama/llama3

# Adjust detail level
thom summarize 2301.00001 --detail brief
thom summarize 2301.00001 --detail detailed

# Output in different languages
thom summarize 2301.00001 --language french

# JSON output
thom summarize 2301.00001 --json

# List supported models
thom models

Configuration

API Keys

Set your API key via environment variables:

# OpenAI
export OPENAI_API_KEY="sk-..."

# Anthropic
export ANTHROPIC_API_KEY="sk-ant-..."

# Or set programmatically
import thom
thom.set_api_key("openai", "sk-...")

Using Local Models with Ollama

# First, install and run Ollama
ollama run llama3

# Then use with thom
thom summarize 2301.00001 --model ollama/llama3

import thom

paper = thom.fetch_paper("2301.00001")
summary = thom.summarize(paper, model="ollama/llama3")

API Reference

Core Functions

thom.fetch_paper(identifier)

Fetch a paper from arXiv by ID or URL.

paper = thom.fetch_paper("2301.00001")
paper = thom.fetch_paper("https://arxiv.org/abs/2301.00001")
paper = thom.fetch_paper("https://arxiv.org/pdf/2301.00001.pdf")

Returns an ArxivPaper object with:

• arxiv_id: The arXiv identifier
• title: Paper title
• authors: List of author names
• abstract: Paper abstract
• categories: arXiv categories
• published: Publication date
• pdf_url: URL to PDF
• arxiv_url: URL to arXiv page

thom.summarize(paper, model="gpt-4o-mini", detail_level="medium", language="english")

Generate a summary of a paper.

summary = thom.summarize(paper)
summary = thom.summarize(paper, model="gpt-4o", detail_level="detailed")
summary = thom.summarize(paper, language="spanish")

Returns a Summary object with:

• paper: The original ArxivPaper
• summary: Generated summary text
• key_points: List of key points
• model: Model used for summarization

thom.search(query, max_results=10, sort_by="relevance")

Search for papers on arXiv.

papers = thom.search("machine learning")
papers = thom.search("au:hinton", max_results=20)
papers = thom.search("cat:cs.LG", sort_by="submittedDate")

thom.compare(papers, model="gpt-4o-mini")

Generate a comparative analysis of multiple papers.

papers = [thom.fetch_paper(id) for id in ids]
analysis = thom.compare(papers)

Supported Models

thom uses LiteLLM for LLM support, which means you can use models from:

• OpenAI: gpt-4o, gpt-4o-mini, gpt-4-turbo, gpt-3.5-turbo
• Anthropic: claude-3-5-sonnet-20241022, claude-3-opus-20240229, claude-3-sonnet-20240229
• Google: gemini/gemini-1.5-pro, gemini/gemini-1.5-flash
• Cohere: command-r-plus, command-r
• Ollama (local): ollama/llama3, ollama/mistral, ollama/mixtral
• Together AI: together_ai/meta-llama/Llama-3-70b-chat-hf
• And many more

List available models:

models = thom.list_supported_models()

License

MIT License