Paper Finder Skill

Purpose

Find academic paper metadata and associated source code repositories from multiple data sources. This is the first step in our paper reproduction pipeline.

When to Use

• User asks to "find", "search", or "look up" a paper
• User provides an arXiv ID, DOI, paper title, or URL
• User asks to "reproduce" or "replicate" a paper (run this first to find the paper and code)
• User wants to know if a paper has official source code

How to Use

Step 1: Run the Search Script

Execute the Python script with the user's query:

python skills/paper-finder/scripts/search_paper.py "<query>" --output workspace/paper_info.json

Supported query formats:

• arXiv ID: 1706.03762, 2301.12345
• arXiv URL: https://arxiv.org/abs/1706.03762
• DOI: 10.5555/3295222.3295349
• Paper title: Attention Is All You Need
• PDF path: /path/to/paper.pdf (uses filename as title hint)

Step 2: Read the Results

After execution, read workspace/paper_info.json to get:

{
  "found": true/false,
  "paper": {
    "title": "...",
    "authors": ["..."],
    "arxiv_id": "...",
    "doi": "...",
    "pdf_url": "...",
    "venue": "...",
    "citation_count": 123,
    "journal_url": "https://doi.org/..."
  },
  "code": {
    "found": true/false,
    "repositories": [
      {
        "url": "https://github.com/...",
        "is_official": true/false,
        "confidence": 0.85,
        "source": "hf_papers|abstract_url|github_search",
        "reason": "..."
      }
    ]
  }
}

Step 3: Interpret Results

• is_official: true (confidence ≥ 0.50): Likely the authors' official repository
• source: "abstract_url": GitHub URL was found directly in the paper text — very reliable
• source: "hf_papers": Repository linked by HuggingFace community — reliable
• source: "github_search": Found via GitHub search — verify manually

Step 4: Present Findings to User

Summarize the results in a clear format:

1. Paper title, authors, year
2. PDF download link
3. Code repositories found (sorted by confidence)
4. Whether official code was identified

Dependencies

• Python 3.10+
• requests library (pip install requests)

Optional: GitHub Token

Set GITHUB_TOKEN environment variable for higher GitHub API rate limits (30 req/min vs 10 req/min). Without a token, the tool still works but may hit rate limits during heavy use.

Error Handling

• If Semantic Scholar returns 429 (rate limited), the search continues with other sources
• If no paper is found, suggest the user try a different query format
• If paper is found but no code, inform the user that the paper may not have public code
• Check search_log in the output for detailed step-by-step information

Data Sources

1. arXiv API — Paper metadata and PDF links
2. Semantic Scholar — Journal DOI, citation count, cross-references
3. HuggingFace Papers — Paper → GitHub repo mapping (replaced Papers with Code)
4. GitHub Search — Fallback repository search