Multi-Model Research (MMR)

Orchestrate deep research across multiple AI models to produce comprehensive, professional-grade research documents.

Goal

Produce research output comparable to or exceeding OpenAI's o3 Deep Research - comprehensive, insightful, PhD-level analysis with no information loss.

Architecture

Optimized for Claude Code Max plan - Claude Code (your session) acts as the orchestrator for FREE.

┌─────────────────────────────────────────────────────────────────────────┐
│                         Claude Code (FREE)                               │
│                         Your session = Orchestrator                      │
│                                                                          │
│   1. Plan research angles                                                │
│   2. Dispatch parallel researchers                                       │
│   3. Validate via subagents (FREE)                                       │
│   4. Aggregate via Gemini 1M                                             │
└─────────────────────────────────────────────────────────────────────────┘
                                    │
                    ┌───────────────┼───────────────┐
                    ▼               ▼               ▼
          ┌─────────────┐ ┌─────────────┐ ┌─────────────┐
          │ Researcher 1│ │ Researcher 2│ │ Researcher 3│  PARALLEL
          │ (Perplexity)│ │ (Perplexity)│ │ (Perplexity)│
          │ Sonar Pro   │ │ Sonar Pro   │ │ Sonar Pro   │  Built-in
          └──────┬──────┘ └──────┬──────┘ └──────┬──────┘  web search
                 │               │               │
                 └───────────────┼───────────────┘
                                 ▼
          ┌─────────────────────────────────────────────┐
          │           Claude Code Subagents             │  PARALLEL
          │           (FREE validation)                 │
          └─────────────────────────────────────────────┘
                                 │
                                 ▼
          ┌─────────────────────────────────────────────┐
          │             Gemini Pro 3                    │
          │             1M context aggregation          │
          └─────────────────────────────────────────────┘

Cost Analysis

Component	Provider	Cost per Run
Orchestration	Claude Code	$0 (Max plan)
Research (6 angles)	Perplexity Sonar Pro	~$0.30
Validation (6 angles)	Claude Code subagents	$0 (Max plan)
Aggregation	Gemini Pro 3	~$0.05
Total		~$0.35

Compare to full API orchestration: ~$5+

Installation

cd ~/Documents/repos/multi-model-research
uv sync

Configuration

Set API keys:

export PERPLEXITY_API_KEY="pplx-..."
export GOOGLE_API_KEY="..."  # For Gemini aggregation

Usage from Claude Code

The recommended way to use MMR is directly from Claude Code, which acts as your orchestrator:

from multi_model_research.library import (
    research_angles_parallel,
    aggregate_research,
    save_raw_research,
)

# Define research angles
angles = [
    {"id": "legal", "title": "Legal Framework", "questions": ["What laws govern X?"]},
    {"id": "market", "title": "Market Analysis", "questions": ["What are current trends?"]},
    # ... more angles
]

# Run parallel research (Perplexity Sonar Pro with web search)
results = research_angles_parallel(angles, instructions="Focus on 2024-2026 data")

# Save raw research for reference
save_raw_research(results, "/path/to/raw.json")

# Aggregate into final report (Gemini 1M)
report = aggregate_research(
    topic="Your Research Topic",
    research_outputs=results,
    output_path="/path/to/final-report.md"
)

CLI Usage

# Basic research
mmr research "AI safety approaches" ./output/report.md

# With options
mmr research "Mexico real estate" ~/Documents/report.md \
    --agents 4 \
    --context "Focus on Baja California" \
    --preset perplexity-research

Presets

Preset	Orchestrator	Researchers	Best For
perplexity-research	Claude Code	Perplexity Sonar Pro	Current web data
opus-gemini	Claude Opus	Gemini Pro 3	Deep analysis
all-gemini	Gemini Pro 3	Gemini Pro 3	1M context

Key Design Principles

1. Claude Code as orchestrator - Leverage your Max plan, no API costs for planning
2. Perplexity for research - Built-in web search with citations
3. Parallel execution - Research all angles simultaneously
4. Subagent validation - Free validation via Claude Code subagents
5. No premature summarization - Research outputs are FULL content
6. Aggregation not summarization - Final report preserves all details

Development Status

See DEVELOPMENT.md for detailed iteration history and current challenges.

What Works

• Parallel research execution (~2 min for 6 angles)
• Perplexity web search integration
• Claude Code subagent validation
• Raw research export (JSON)

Current Challenge

• LLM aggregation tends to compress/summarize
• Exploring mechanical compilation + LLM analysis as separate steps

Pydantic Schemas

• ResearchRequest - Input configuration
• ResearchAngle - A specific research area
• ResearchOutput - Full research from one agent
• ValidationResult - Pass/fail with issues
• FinalReport - Aggregated output

License

MIT