multi-model-research（多模型研究）

Multi-Model Research Agent

Implements Karpathy's LLM Council pattern for superior research through parallel queries, peer review, and chairman synthesis.

Architecture

Geoffrey/Claude (Native Council Member):

• Routes simple vs complex queries
• Calls external API orchestrator (research.py)
• Provides my own research response
• Conducts peer review phase
• Requests GPT-5.1 synthesis (chairman)
• Saves final report to Obsidian

Python External API Orchestrator:

• Fetches responses from GPT-5.1, Gemini 3.0 Pro, Perplexity Sonar, Grok 4.1
• Returns JSON with all external responses
• I handle all orchestration and synthesis

When to Use This Skill

Use multi-model research when:

• Complex analysis needed - Multiple perspectives valuable
• Factual verification critical - Cross-model validation
• Comprehensive coverage required - No single model sufficient
• Current information essential - Perplexity provides web grounding
• Contested topics - Benefit from diverse model perspectives

Simple vs Council Mode

Simple Mode (Perplexity only):

• Factual lookups
• Current events
• Quick research with citations
• Completes in <15 seconds

Council Mode (Full council):

• Comparative analysis
• Deep research
• Multiple perspectives needed
• Strategic questions
• Completes in <90 seconds

Workflow

Simple Query

User: "What are the latest developments in quantum computing?"
     ↓
I decide: Simple query (factual, current)
     ↓
I call: uv run scripts/research.py --query "..." --models perplexity
     ↓
I read: JSON response from Perplexity
     ↓
I format: Markdown report with citations
     ↓
I save: To Obsidian Geoffrey/Research folder
     ↓
I return: Summary to user with Obsidian link

Council Query

User: "Compare the AI strategies of OpenAI, Anthropic, and Google"
     ↓
I decide: Council query (comparative, complex)
     ↓
I call: uv run scripts/research.py --query "..." --models gpt,gemini,perplexity,grok
     ↓
I read: JSON with all external responses
     ↓
I provide: My own (Claude) research response
     ↓
I conduct: Peer review (each model ranks others)
     ↓
I request: GPT-5.1 chairman synthesis
     ↓
I format: Comprehensive markdown report
     ↓
I save: To Obsidian Geoffrey/Research folder
     ↓
I return: Summary with Obsidian link

Output Format

All research reports saved to Obsidian include:

• Executive Summary (2-3 paragraphs)
• Key Findings (organized by theme, inline citations)
• Confidence Assessment (what's certain vs debated)
• References Section (all sources with URLs and dates)

Citations use numeric format: [1], [2], etc.

Technical Details

Python Script:

cd skills/multi-model-research
uv run scripts/research.py --query "Your question" --models perplexity --output /tmp/responses.json

Config:

• config.yaml - Model settings, routing rules
• prompts/system_prompts.yaml - Per-model system prompts
• prompts/peer_review.md - Peer review template
• prompts/chairman_synthesis.md - GPT-5.1 synthesis template

Dependencies:

• httpx (async HTTP client)
• pyyaml (config parsing)
• python-dotenv (env vars)
• python-frontmatter (Obsidian frontmatter)

API Keys Required:

• OPENAI_API_KEY (GPT-5.1)
• GEMINI_API_KEY (Gemini 3.0 Pro)
• PERPLEXITY_API_KEY (Sonar Pro)
• XAI_API_KEY (Grok 4.1)

All keys configured in ~/.env file.

Examples

Simple Research:

User: "What is RAG in AI?"

I route to: Simple mode (Perplexity)
Output: Concise explanation with current examples and citations
Time: ~10 seconds

Council Research:

User: "Compare serverless vs containers for production ML workloads"

I route to: Council mode (all 4 external + me)
Process:
  1. GPT-5.1: Provides comprehensive technical comparison
  2. Gemini 3.0: Analyzes cost and performance trade-offs
  3. Perplexity: Current industry trends and case studies
  4. Grok 4.1: Developer sentiment from X/Twitter
  5. Claude (me): Synthesize with nuanced analysis
  6. Peer review: Each model ranks others
  7. GPT-5.1 (chairman): Final synthesis

Output: Multi-perspective analysis with citations
Time: ~60 seconds

Limitations

• Cost: Council mode uses 4-5 API calls per query
• Latency: Council mode takes 60-90 seconds
• API Limits: Rate limits may throttle parallel requests
• Citation Quality: Non-Perplexity models require URL extraction

Future Enhancements

• Streaming responses during deliberation
• Cost tracking and budget limits
• Query history and versioning
• Custom model weights based on topic
• Integration with Geoffrey's knowledge base

*This skill implements Karpathy's LLM Council pattern released November 22, 2025.*