arxiv-ingest 0.2.0

Fetch recent arXiv papers and scaffold them into a structured research wiki

arxiv-ingest

Fetch recent arXiv papers by keyword and scaffold them into structured research notes.

Works standalone as a CLI, or integrates with Claude Code for LLM-assisted evidence extraction and wiki generation.

Features

• Keyword-based collection — list topics in config.yaml, no code needed
• 3-layer output — sources/ (metadata) · evidence/ (claims) · wiki/ (synthesis)
• Safe re-runs — already-filled files are never overwritten
• GitHub Actions template — auto-runs Mon–Fri, workflow_dispatch for on-demand runs
• Any wiki — point output_dir at any directory

Quick start

# Install (requires Python 3.11+)
git clone https://github.com/yamadan96/arxiv-ingest
cd arxiv-ingest
uv sync                         # or: pip install -e .

# Configure
cp config.yaml.example config.yaml
# Edit config.yaml: set keywords, output_dir, and category mappings

# Fetch papers and generate skeletons
arxiv-ingest run                # fetch + generate in one step

# Or step by step:
arxiv-ingest fetch              # saves data/fetched.json
arxiv-ingest generate           # creates skeleton files in output_dir

Configuration (config.yaml)

output_dir: "../research-wiki"  # path to your wiki directory
max_results: 20                 # papers per keyword
days_back: 7                    # look back N days

keywords:
  - "Vision-Language Model transformer"
  - "LoRA PEFT fine-tuning language model"
  - "LLM reasoning chain-of-thought"

# Only accept papers in these arXiv categories:
allowed_arxiv_categories: [cs.CV, cs.CL, cs.LG, cs.AI, cs.RO, stat.ML]

# Primary category must be in this list (prevents off-topic papers):
require_primary_in: [cs.CV, cs.CL, cs.AI, cs.RO, stat.ML]

# arXiv category → wiki folder name:
category_map:
  cs.CV: "Multimodal"
  cs.CL: "Post_Training"
  cs.AI: "Reasoning"
  cs.RO: "Physical_AI"
  stat.ML: "Pretraining"

See config.yaml.example for the full reference with all options documented.

Output format

Skeleton files are created at:

output_dir/
├── sources/{Category}/{slug}.md      # Metadata (title, authors, abstract, URL)
├── evidence/{Category}/{slug}.md     # Claims & benchmarks — fill this in
└── wiki/papers/{Category}/{slug}.md  # Synthesis & interpretation — fill this in

sources/ files are written once and never touched again. evidence/ and wiki/ files are only (re-)created when they still contain the unfilled template placeholder, so your edits are safe across re-runs.

Using with Claude Code

Copy the skill file so /arxiv-ingest is available as a Claude Code command:

cp .claude/commands/arxiv-ingest.md ~/.claude/commands/
# or into your project's .claude/commands/

After running arxiv-ingest run, invoke /arxiv-ingest in Claude Code. Claude reads each paper's abstract (and PDF when available) and fills in the evidence/ and wiki/ files automatically.

GitHub Actions: daily auto-ingest

Setup

1. Fork or use this repo as a template

2. Create a GitHub Fine-grained PAT with Write access to your wiki repo

3. Add to your repo's Settings → Secrets and variables → Actions:

   Kind	Name	Value
   Secret	GH_PAT	your PAT
   Variable	WIKI_REPO	your-name/research-wiki

4. Go to Actions → Daily arXiv Ingest → Run workflow to verify

Schedule

• Auto: Mon–Fri UTC 01:00 (JST 10:00) — cron: '0 1 * * 1-5'
• Manual: workflow_dispatch with optional days_back override

Requirements

• Python 3.11+
• uv (recommended) or pip

License

MIT