IDENTITY.md — GraphRAG Agent Configuration
Agent Name
GraphRAG-Agent
Version
1.0.0
Architecture
AI Factory Model — 4 separated layers:
- Graph Layer: TigerGraph Cloud — entity storage, multi-hop traversal, GSQL queries
- Orchestration Layer: Dual pipeline routing with adaptive complexity analysis
- LLM Layer: Universal provider (12+ supported) — generation, extraction, analysis
- Evaluation Layer: RAGAS + custom F1/EM metrics + cost tracking
Supported LLM Providers
| Provider | API Key Env Var | Default Model |
|---|---|---|
| OpenAI | OPENAI_API_KEY |
gpt-4o-mini |
| Anthropic Claude | ANTHROPIC_API_KEY |
claude-sonnet-4 |
| Google Gemini | GEMINI_API_KEY |
gemini-2.0-flash |
| Mistral AI | MISTRAL_API_KEY |
mistral-large-latest |
| Cohere | COHERE_API_KEY |
command-r-plus |
| Ollama (Local) | — | llama3.2 |
| OpenRouter | OPENROUTER_API_KEY |
llama-3.3-70b |
| Groq | GROQ_API_KEY |
llama-3.3-70b-versatile |
| xAI Grok | XAI_API_KEY |
grok-3-mini |
| Together AI | TOGETHER_API_KEY |
llama-3.1-70b-turbo |
| HuggingFace | HF_TOKEN |
llama-3.3-70b |
| DeepSeek | DEEPSEEK_API_KEY |
deepseek-chat |
Graph Database
- Engine: TigerGraph Cloud (free tier compatible)
- Schema: Document → Chunk → Entity → Community
- Edges: PART_OF, MENTIONS, RELATED_TO, IN_COMMUNITY
- Queries: vectorSearchChunks, vectorSearchEntities, graphRAGTraverse
Evaluation Dataset
- Primary: HotpotQA (distractor setting, 90K+ questions)
- Metrics: F1, Exact Match, Context Hit Rate, Faithfulness, Answer Relevancy
Communication Channels
- Web Dashboard (Next.js): http://localhost:3000
- Python CLI:
python -m graphrag.main - OpenClaw Skills API: local HTTP endpoints
- Telegram/Discord: via OpenClaw messaging integration
Security
- All API keys stored as environment variables (never in code)
- TigerGraph auth via token-based authentication
- OpenClaw ClawKeeper security policies enforced
- No arbitrary code execution on host