arDive: a simple dive into your ArXiv

A small command-line agent that pulls papers from arXiv and uses Llama3.2 to summarize, explain, compare, and digest them. Anyone can easily install and use without the need of a paid plan.

Install

# 1. Install Ollama (free, runs models locally): https://ollama.com
ollama pull llama3.2:1b     # the fast default; or any open model you like

# 2. Install arDive
pip install ardive

That's it. arDive talks to the local Ollama server. Pick a different model with ARDIVE_MODEL (e.g. export ARDIVE_MODEL=qwen2.5), or point at a remote Ollama with OLLAMA_HOST.

From source

git clone https://github.com/rohankosalge/arDive
cd arDive
pip install -e .

Usage

# Summarize a paper as bullet points
ardive sum 1234.56789

# Focus on one section, cap the bullets
ardive sum 1234.56789 --section methodology --max-bullets 5

# Explain like I'm 5 (works on every command)
ardive sum 1234.56789 --eli5

# Compare papers (similarities + a differences table)
ardive comp 1234.56789 9876.54321

# List papers on a topic (fast; no model)
ardive dig "diffusion models for protein folding"
ardive dig "graph neural networks" -n 5

# List papers similar to a given one (fast; no model)
ardive sim 1234.56789 -n 5

Commands

Flags

• --eli5 — explain in plain, jargon-free language (sum, comp).
• --section {abstract,intro,methodology,related,citations} — sum only; focus on one section.
• --max-bullets N — sum only; hard cap on the number of bullets (positive integer).
• -n/--num N — dig and sim; how many papers to list (default 3).

How it works

sum and comp download each paper's PDF and extract its full text, then send it to a local open-source model via Ollama (default llama3.2:1b) with a prompt tailored to the command. dig and sim use no model — they're pure arXiv lookups that just list papers, so they're fast. In a terminal the response is rendered as formatted Markdown inside a box; when piped or redirected (e.g. ardive sum 1234.56789 > out.md) it's written as plain Markdown so the file stays clean.

Long papers can exceed the model's context window and be truncated. arDive asks Ollama for an 8192-token window by default; raise it (at the cost of more RAM) with export ARDIVE_NUM_CTX=16384.

Speed

Summaries run entirely on your machine, so wall-clock time is dominated by the model. A few levers:

• Model choice is the biggest one. The default llama3.2:1b is fast but modest in quality; for better summaries (slower) try export ARDIVE_MODEL=llama3.2 (~3B) or export ARDIVE_MODEL=qwen2.5:3b. 7B+ models are noticeably slower on full papers.
• Abstract is near-instant. ardive sum <id> --section abstract skips the PDF download and summarizes just the abstract.
• First run is slowest. It loads the model into memory; arDive keeps it warm for 15 min afterward (tune with ARDIVE_KEEP_ALIVE), so repeat runs are quicker.
• Smaller asks finish sooner. --max-bullets N shortens the output, and a lower ARDIVE_NUM_CTX trades context for speed.