academic-talon（学术爪子）

🎓 Academic Talon Skill

Your AI-powered academic research assistant for paper search → BibTeX extraction → Zotero archiving → local PDF serving.

Save hours of manual work searching papers, copying citations, and organizing your library.

🎯 What it does (when to use this skill)

Trigger this skill when the user wants to:

🔧 Architecture & Dependencies

This is a toolbox skill that provides multiple independent academic research tools. You can use just the features you need. A common complete workflow looks like this:

User Query
    ↓
[academic-talon] ← this skill
    ↓
1. Search → Multiple search APIs (arXiv, Google Scholar via SerpAPI, etc.)
    ↓
2. PDF Download → saved to local `pdfs/` directory
    ↓
3. PDF Parsing → **GROBID service** processes PDF
    ↓
   - Header analysis → extracts metadata → skill generates clean BibTeX
   - Full text analysis → returns complete TEI XML with full document structure
    ↓
4. If header analysis: BibTeX Generation → skill formats clean publication-ready output
    ↓
5. Zotero Archiving → via **pyzotero** → your Zotero library → auto-add to collection
    ↓
6. PDF Serving → built-in HTTP server serves PDFs from your intranet
    ↓
Result: Paper in Zotero with working PDF link, clean BibTeX ready for citation

*You don't have to use this full workflow - use individual tools as needed.*

Required External Services

⚙️ Setup Instructions

1. Install Python dependencies

pip install -r skills/academic-talon/requirements.txt

2. Configure environment variables (skills/academic-talon/.env)

# ========== Zotero Configuration (Required for archiving) ==========
ZOTERO_API_KEY=your_zotero_api_key_here
ZOTERO_LIBRARY_ID=your_library_id_here
ZOTERO_LIBRARY_TYPE=user  # or "group" for group libraries

# ========== GROBID Configuration (Required for PDF parsing) ==========
GROBID_API_URL=http://localhost:8070/api
# Or if you use Docker Compose behind nginx:
# GROBID_API_URL=http://localhost:8080/api

# ========== Optional Search API Keys ==========
# Get these from their respective websites
SEMANTIC_SCHOLAR_API_KEY=your_semantic_scholar_api_key
SERPAPI_KEY=your_serpapi_key_for_google_scholar
TAVILY_API_KEY=your_tavily_api_key

# ========== Local PDF Serving (Optional) ==========
# After starting the PDF server, set this to your intranet URL:
# Example: PDF_BASE_URL=http://192.168.1.100:8000/
PDF_BASE_URL=http://your-server-ip:port/

3. Start GROBID (for PDF parsing)

Option A: Docker Compose (Recommended)

Create compose.yml in your GROBID directory:

version: "3.9"
services:
  grobid:
    # Choose the right image for your hardware:
    # - For non-GPU environments: grobid/grobid:0.8.2-crf (CRF-only model, smaller)
    # - For GPU environments: grobid/grobid:0.8.2-full (includes CRF + deep learning models)
    image: grobid/grobid:0.8.2-crf
    container_name: grobid
    restart: unless-stopped
    expose:
      - "8070"
    environment:
      JAVA_OPTS: "-Xms512m -Xmx4g"
    volumes:
      - ./grobid/tmp:/opt/grobid/tmp
      - ./grobid/logs:/opt/grobid/logs

💡 Image selection: Use grobid/grobid:0.8.2-crf for CPU-only / non-GPU environments (smaller image, faster startup). Use grobid/grobid:0.8.2-full if you have GPU and want maximum accuracy with deep learning models.

Start:

docker-compose up -d

Option B: Direct run

Follow GROBID documentation to run directly.

4. (Optional) Start the Local PDF Server

If you want to serve downloaded PDFs locally:

# Start on port 8000, allow all intranet access
python skills/academic-talon/scripts/start_pdf_server.py start 8000 内网

# Check status
python skills/academic-talon/scripts/start_pdf_server.py status

# Stop
python skills/academic-talon/scripts/start_pdf_server.py stop

The server:

• Serves only from the pdfs/ directory (sandboxed, no access outside)
• Default binds to all interfaces → accessible from your entire intranet
• Filenames are citation keys (e.g. zhang2025hallucinationdetection.pdf)
• When PDF_BASE_URL is configured, archived papers automatically get the correct local URL

📖 Usage (for LLM)

Input Schema

Output Format

All actions return JSON in this format:

{
  "success": true,
  "action": "search",
  "query": "your search query",
  "results": [
    {
      "title": "Paper Title",
      "authors": ["Author One", "Author Two"],
      "year": "2025",
      "abstract": "Paper abstract...",
      "url": "https://...",
      "pdf_url": "https://...",
      "source": "arxiv"
    }
  ]
}

✨ Features (and how they help your research)

1. Fixed arXiv Search

• ❌ Before: arXiv API defaults to OR semantics → searching "LLM judge knowledge possession" returns papers with just one keyword → many irrelevant results
• ✅ Now: Proper AND semantics matches what you get in browser search. Every result contains all query terms in title or abstract.
• 🎯 Benefit: Get relevant results first try, no scrolling through irrelevant papers

2. Publication-Ready BibTeX Generation

• Follows exactly the format used by top AI conferences (NeurIPS, ICML, ICLR, CVPR, etc.)
• Correct entry types:

- Journal article → @article - Conference paper → @inproceedings with conference name in booktitle - arXiv preprint → @article with journal = {arXiv preprint xxxx.xxxxx} exactly matching your example

• Cleans up junk: removes unnecessary fields like date, month, publisher, day that shouldn't be in final submissions
• Correct citation keys: lastnameYearTitle → zhang2025hallucinationdetection matches standard academic practice

Example output (ready to paste into your manuscript):

@article{zhang2025hallucinationdetection,
  author = {Zhang, Chenggong and Wang, Haopeng},
  title = {Hallucination Detection and Evaluation of Large Language Model},
  year = {2025},
  journal = {arXiv preprint 2512.22416},
  abstract = {Hallucinations in Large Language Models...},
}

@inproceedings{gal2016dropout,
  author = {Gal, Yarin and Ghahramani, Zoubin},
  title = {Dropout as a bayesian approximation: Representing model uncertainty in deep learning},
  booktitle = {ICML},
  year = {2016},
}

3. Smart Zotero Archiving

• 🎯 Default collection: all papers go to openclaw unless you specify otherwise
• 🪄 Auto-creation: if the collection doesn't exist, skill automatically creates it
• 🔄 Smart duplicate handling: if paper already exists in your library, skill adds it to the target collection instead of failing
• 🏷️ Correct Zotero types: preprint → preprint, conference → conferencePaper, journal → journalArticle
• 📍 Local PDF links: when you run the local PDF server, links point directly to your local copy

Benefit: Build your research library without repetitive manual clicking.

4. Local PDF Library Serving

• Maintain all your PDFs locally
• Built-in HTTP server with start/stop/status management
• Designed for intranet access → you can access your PDFs from any device on your network
• Zotero links point directly to local files → no downloading the same PDF multiple times

🔒 Security Considerations

⚠️ Important Security Notes

1. PDF Processing goes to GROBID:

- This skill sends PDF content to the configured GROBID_API_URL for metadata extraction - Recommendation: Run GROBID locally on your own machine/infrastructure for privacy - If you use a third-party GROBID service, be aware that they will see your PDFs

1. Local PDF Server:

- This skill runs an HTTP server that serves PDF files from the pdfs/ directory - It is designed for intranet/private network use only - The server does NOT include authentication - ❌ Do NOT expose this server directly to the public internet - ✅ Only run on trusted private networks, or put it behind a reverse proxy with authentication

1. File Access Restrictions:

- All file operations (download, analysis) are sandboxed to the pdfs/ directory within this skill's installation - Directory traversal attacks are prevented by path checking - The skill cannot access or modify files outside its own directory

1. API Key Storage:

- All API keys are stored locally in the .env file - Never commit .env to version control - Keys are only used for API requests directly from your machine to the service providers

Best Security Practices

• ✅ Run GROBID locally (don't send sensitive PDFs to third parties)
• ✅ Keep PDF server on private/intranet network only
• ✅ Use reverse proxy with authentication if you need public access
• ✅ Use a dedicated Zotero API key with limited permissions
• ✅ Don't expose GROBID directly to the internet (use the recommended nginx proxy with IP whitelist)

📋 Complete Workflow Example

# 1. Search for papers
result = skill.run({
  "action": "search",
  "query": "LLM judge knowledge possession",
  "limit": 5
})

# 2. Download PDF for first result
paper = result["results"][0]
download_result = skill.run({
  "action": "download",
  "url": paper["pdf_url"],
  "paper_info": paper
})

# 3. Extract BibTeX from downloaded PDF
analyze_result = skill.run({
  "action": "analyze",
  "pdf_input": download_result["pdf_path"],
  "analysis_type": "header"
})

# 4. Archive to Zotero (goes to openclaw collection by default)
paper["bibtex"] = analyze_result["result"]
archive_result = skill.run({
  "action": "archive",
  "paper_info": paper
})

if archive_result["success"]:
  print(f"✅ Paper archived to Zotero: {archive_result['result']['item_id']}")

🐛 Troubleshooting

📊 Benefits for Academic Research

• Saves time: Go from keywords → archived paper in minutes instead of manually copying everything
• Consistent citations: Always get clean BibTeX ready for journal/conference submission
• Organized library: Automatic collection management keeps your papers organized
• Local access: Keep all PDFs locally and access them from anywhere on your network
• Correct search: Get relevant results from arXiv with proper AND semantics

📦 Dependencies Summary

• Python: 3.6+
• Python packages: requests, python-dotenv, pyzotero
• External services: GROBID (PDF parsing), Zotero API (archiving)
• Optional APIs: SerpAPI (Google Scholar), Semantic Scholar API, Tavily API

📄 License

MIT License - free for academic and commercial use.