Fetch ArXiv Papers

Fetch and store AI/ML research papers from arXiv with deduplication.

When to Use

Use this skill when you need to:

• Collect recent AI/ML research papers from arXiv
• Store papers in memory for later analysis
• Avoid processing duplicate papers

Instructions

Step 1: Define ArXiv Categories

Target these arXiv categories for AI/ML papers:

• cs.AI - Artificial Intelligence
• cs.LG - Machine Learning
• cs.CL - Computation and Language (NLP)
• cs.CV - Computer Vision
• cs.NE - Neural and Evolutionary Computing

Step 2: Fetch ArXiv RSS Feeds

Use the rss tool to fetch papers from each category feed.

Feed URL pattern:

• https://export.arxiv.org/rss/{category}
• Example: https://export.arxiv.org/rss/cs.AI

For each feed:

1. Call the rss tool with the feed URL
2. Extract: title, link (abstract URL), description (abstract), dc:creator (authors), published date
3. Parse the arXiv ID from the link URL (e.g., 2401.12345 from https://arxiv.org/abs/2401.12345)

Step 3: Check for Duplicates

For each paper from the feeds:

1. Check if already seen:

   • Call memory/check_seen with key=arXiv ID, namespace="news/papers"
   • If seen=true, skip this paper

2. Validate required fields:

   • Paper must have: title, arXiv ID, abstract
   • Skip papers missing required fields

Step 4: Store New Papers

For each new (unseen) paper:

1. Store in memory:

   • Call memory/add with:
     • type: "document"
     • namespace: "news/papers"
     • data: {arxiv_id, title, authors, abstract, categories, published_date, pdf_url, abstract_url}
     • metadata: {fetched_at, source_category}

2. Mark as seen:

   • Call memory/mark_seen with:
     • key: arXiv ID
     • namespace: "news/papers"
     • ttl_seconds: 2592000 (30 days)

Step 5: Return Results

Return a summary including:

• Number of papers stored
• Number of duplicates skipped
• Number of categories processed
• Any failed feeds

Tool Usage Guidance

rss tool

• Use to fetch arXiv RSS feed content
• Handles XML parsing and entry extraction
• Returns list of entries with title, link, description, published

memory/check_seen

• Call before processing each paper
• Key should be the arXiv ID (e.g., "2401.12345")
• Returns {seen: true/false}

memory/add

• Store each new paper
• Include all extracted metadata
• Type should be "document"

memory/mark_seen

• Call after successfully storing
• Use 30-day TTL for papers (longer than news articles)

Paper Data Schema

{
  "arxiv_id": "2401.12345",
  "title": "Advances in Large Language Model Reasoning",
  "authors": ["Alice Smith", "Bob Jones"],
  "abstract": "We present a novel approach to...",
  "categories": ["cs.AI", "cs.CL"],
  "published_date": "2026-01-25",
  "pdf_url": "https://arxiv.org/pdf/2401.12345.pdf",
  "abstract_url": "https://arxiv.org/abs/2401.12345"
}

Error Handling

• If a feed fails to fetch, log the error and continue with other feeds
• If memory operations fail, log but don't crash
• Return partial results if some feeds succeed

Success Criteria

• At least one category feed successfully fetched
• New papers stored in memory
• Duplicates correctly identified and skipped by arXiv ID
• Failed feeds logged but don't stop collection