AI-powered research automation using LangGraph orchestration, multiple search engines, and iterative verification. Built with hexagonal architecture for enterprise-grade research workflows.
Enable it via Settings → Pages:
- Source: GitHub Actions
CAG (Causal Analysis Graph) Deep Research is a sophisticated research automation platform that combines multiple AI agents, search engines, and verification loops to produce comprehensive, fact-checked research reports. Unlike simple Q&A systems, CAG performs iterative deep research with automatic quality assessment and knowledge graph construction.
- Multi-Agent Architecture: Specialized agents for search, analysis, verification, and reporting
- Iterative Deepening: Automatically identifies knowledge gaps and performs follow-up research
- Dual Search Integration: Combines Tavily and Exa APIs for comprehensive coverage
- Quality Assurance: Built-in audit feedback and verification loops
- Knowledge Graph: Constructs domain entities and relationships during research
- Hexagonal Architecture: Clean separation between domain logic, ports, and adapters
| Category | Technologies |
|---|---|
| Orchestration | LangGraph, LangChain Core |
| Search | Tavily API, Exa API, DuckDuckGo (free) |
| LLM | GitHub Models (free), Groq, DeepSeek, Ollama (local) |
| Architecture | Hexagonal/Ports & Adapters, DDD |
| Data | Pydantic, httpx (async) |
# Set API keys
export TAVILY_API_KEY="your_key"
export EXA_API_KEY="your_key"
# Clone and install
git clone https://github.com/jayhemnani9910/revolu-idea.git
cd revolu-idea
pip install -r requirements.txt
# Run research
python main.py "What are the latest developments in quantum computing?"User Query
↓
[Search Planner] → Plans research strategy
↓
[Web Searcher] → Queries Tavily + Exa
↓
[Content Analyzer] → Extracts key information
↓
[Knowledge Builder] → Constructs entity graph
↓
[Report Generator] → Creates structured report
↓
[Audit Validator] → Verifies facts, checks gaps
↓
Final Report (Markdown)
revolu-idea/
├── domain/ # Core business logic
│ └── entities.py # Research entities
├── ports/ # Interfaces
│ ├── llm_port.py
│ └── search_port.py
├── adapters/ # External integrations
│ ├── ollama_adapter.py
│ ├── tavily_adapter.py
│ └── exa_adapter.py
├── agents/ # LangGraph nodes
│ └── nodes/
├── graph/ # Workflow definition
├── config/ # Settings
└── main.py # CLI entrypoint
Copy .env.example to .env and configure:
# .env file
# LLM Provider (pick one)
# Option 1: GitHub Models (FREE with Copilot subscription)
LLM_PROVIDER=github
LLM_BASE_URL=https://models.inference.ai.azure.com
LLM_MODEL=gpt-4o-mini # or "auto" for model pool
LLM_API_KEY=ghp_xxxxx # GitHub token with models:read scope
# Option 2: Groq (fast, free tier available)
# LLM_PROVIDER=groq
# LLM_BASE_URL=https://api.groq.com/openai/v1
# LLM_MODEL=auto
# LLM_API_KEY=gsk_xxxxx
# Search (DuckDuckGo is free, no key needed)
SEARCH_PROVIDER=duckduckgo
# TAVILY_API_KEY=tvly-xxxxx # for premium search
# Research parameters
MAX_RECURSION_DEPTH=5
MAX_INVESTIGATIONS_PER_EDGE=2| Model | RPM | RPD | Tokens |
|---|---|---|---|
| gpt-4o-mini | 15 | 150 | 8k in, 4k out |
| gpt-4o | 10 | 50 | 8k in, 4k out |
Reports saved to output/reports/ include:
- Executive summary
- Key findings with confidence scores
- Source citations
- Knowledge graph entities
- Verification status
- Follow-up questions
- Academic literature reviews
- Competitive intelligence
- Due diligence and fact-checking
- Technical content research
- Policy and market analysis