🔍 GraphRAG Inference Hackathon — Dual Pipeline System
Proving that graphs make LLM inference faster, cheaper, and smarter — with any LLM provider.
Quick Start · 12 Providers · OpenClaw · Architecture · Benchmarks · Deploy
🚀 Quick Start
Option A: Next.js Dashboard (Recommended)
cd web
npm install
cp .env.example .env.local
# Set ANY provider key — or just use Ollama for free:
npm run dev
# → http://localhost:3000
Option B: Docker (One Command)
docker build -t graphrag .
docker run -p 3000:3000 -e ANTHROPIC_API_KEY=sk-ant-... graphrag
Option C: Python CLI
pip install -r requirements.txt
python -m graphrag.main demo
Option D: Ollama (100% Free, Local)
ollama pull llama3.2
cd web && npm install && npm run dev
# Select "Ollama (Local)" in provider dropdown
🏗️ Architecture (AI Factory Model — 4 Layers)
┌──────────────────────────────────────────────────────────────────────┐
│ LAYER 4: EVALUATION │
│ Next.js Dashboard │ RAGAS │ F1/EM │ Cost Tracking │ Live Benchmark │
├──────────────────────────────────────────────────────────────────────┤
│ LAYER 3: UNIVERSAL LLM (12 Providers) │
│ OpenAI │ Claude │ Gemini │ Mistral │ Ollama │ Groq │ DeepSeek │ … │
├────────────────────────────┬─────────────────────────────────────────┤
│ Pipeline A: Baseline RAG │ Pipeline B: GraphRAG │
│ Query → Vector → LLM │ Query → Keywords → Graph → Context → LLM │
│ │ 🧠 Adaptive Router │ 🔗 Reasoning Paths │
├────────────────────────────┴─────────────────────────────────────────┤
│ LAYER 1: GRAPH (TigerGraph Cloud) │
│ Schema: Document → Chunk → Entity → Community │
│ GSQL: vectorSearchChunks │ vectorSearchEntities │ graphRAGTraverse │
└──────────────────────────────────────────────────────────────────────┘
Each layer is a separate module — swap TigerGraph for Neo4j, Claude for Ollama, or RAGAS for custom evals without touching other layers.
🤖 Supported LLM Providers
| # | Provider | Default Model | Cost/1K tokens | Speed |
|---|---|---|---|---|
| 1 | OpenAI | gpt-4o-mini | $0.00015 in / $0.0006 out | ⚡ Fast |
| 2 | Anthropic Claude | claude-sonnet-4 | $0.003 / $0.015 | 🔵 Medium |
| 3 | Google Gemini | gemini-2.0-flash | $0.0001 / $0.0004 | ⚡ Fast |
| 4 | Mistral AI | mistral-large | $0.002 / $0.006 | 🔵 Medium |
| 5 | Cohere | command-r-plus | $0.0025 / $0.01 | 🔵 Medium |
| 6 | 🦙 Ollama | llama3.2 | $0 / $0 | ⚡ Local |
| 7 | OpenRouter | llama-3.3-70b | $0.0004 / $0.0004 | 🔵 Medium |
| 8 | Groq | llama-3.3-70b | $0.0006 / $0.0008 | ⚡⚡ Blazing |
| 9 | xAI Grok | grok-3-mini | $0.0003 / $0.0005 | ⚡ Fast |
| 10 | Together AI | llama-3.1-70b | $0.0009 / $0.0009 | ⚡ Fast |
| 11 | HuggingFace | llama-3.3-70b | $0 / $0 | 🔵 Medium |
| 12 | DeepSeek | deepseek-chat | $0.00014 / $0.00028 | ⚡ Fast |
How: All providers use OpenAI SDK with dynamic baseURL — zero extra dependencies. Switch providers from the dropdown in the dashboard UI.
🌟 Novel Features
- 🧠 Adaptive Query Router — complexity scoring → auto pipeline selection
- 📋 Schema-Bounded Extraction — 9 entity types + 15 relation types
- 🔑 Dual-Level Keywords — LightRAG-inspired high/low-level retrieval
- 🔗 Graph Reasoning Paths — step-by-step NL traversal explanation
- 🤖 12-Provider Universal LLM — including free Ollama local
- 🦞 OpenClaw Agent Skills — GraphRAG as autonomous agent capabilities
- 📊 Live Benchmark Button — run real evaluations from the dashboard
- 💰 12-Provider Cost Comparison — real-time projections
📊 Benchmarks
Live Benchmark (Run from Dashboard)
Click "🏃 Run Benchmark Now" in the Benchmark tab to evaluate both pipelines on 10 HotpotQA questions with your configured provider. Results populate real-time with F1, EM, token counts, costs.
Expected Results (HotpotQA)
| Metric | Baseline RAG | GraphRAG | Winner |
|---|---|---|---|
| F1 Score | ~0.45–0.60 | ~0.55–0.70 | ✅ GraphRAG |
| Exact Match | ~0.30–0.45 | ~0.35–0.50 | ✅ GraphRAG |
| Tokens/Query | ~800–1000 | ~2000–2800 | ✅ Baseline |
| F1 Win Rate | — | ~55–70% | ✅ GraphRAG |
Key Finding: GraphRAG consistently outperforms baseline on multi-hop questions (bridge type) where connecting facts across documents is required. The token overhead is 2–3×, but the Adaptive Router eliminates this cost for simple queries.
🦞 OpenClaw Integration
Full CIK model (Capability + Identity + Knowledge):
| File | Purpose |
|---|---|
openclaw/SOUL.md |
Agent identity, values, personality |
openclaw/IDENTITY.md |
Configuration, supported providers |
openclaw/MEMORY.md |
Learned facts about GraphRAG |
openclaw/skills/graph_query/ |
NL → knowledge graph traversal |
openclaw/skills/compare_pipelines/ |
Dual-pipeline comparison |
openclaw/skills/cost_estimate/ |
12-provider cost projection |
🧪 Testing
# Run all 31 unit tests
python tests/test_core.py
# Tests cover:
# - cosine_similarity (5 cases including edge cases)
# - chunk_text (4 cases: basic, empty, short, overlap)
# - entity ID generation (3 cases: deterministic, case-insensitive, type-different)
# - F1/EM computation (5 cases: perfect, partial, no overlap, empty)
# - context hit rate (2 cases)
# - token efficiency (3 cases)
# - provider registry (4 cases: completeness, fields, ollama free, available)
# - evaluation layer aggregate + report (2 cases)
🐳 Deployment
Docker
docker build -t graphrag .
docker run -p 3000:3000 \
-e ANTHROPIC_API_KEY=sk-ant-... \
-e OPENAI_API_KEY=sk-... \
graphrag
Vercel
cd web
npx vercel --prod
Env Variables
# Set any/all — system auto-detects available providers
ANTHROPIC_API_KEY=sk-ant-... # Claude
OPENAI_API_KEY=sk-... # GPT-4o
GEMINI_API_KEY=AIza... # Gemini
GROQ_API_KEY=gsk_... # Groq (ultra-fast)
DEEPSEEK_API_KEY=sk-... # DeepSeek (cheapest)
# Or: ollama pull llama3.2 # Free, local
📁 Project Structure (68 files, 240KB)
├── web/ # Next.js 15 Dashboard
│ ├── src/app/
│ │ ├── globals.css # 14KB fused TigerGraph×Claude design system
│ │ └── api/
│ │ ├── compare/route.ts # Multi-provider dual-pipeline API
│ │ ├── benchmark/route.ts # Live benchmark runner with F1/EM
│ │ └── providers/route.ts # Available providers + Ollama health
│ ├── src/components/tabs/
│ │ ├── LiveCompare.tsx # Provider selector + side-by-side comparison
│ │ ├── Benchmark.tsx # Live "Run Now" + radar/bar charts
│ │ ├── CostAnalysis.tsx # 12-provider cost projections
│ │ └── GraphExplorer.tsx # Interactive SVG knowledge graph
│ └── src/lib/
│ ├── llm-providers.ts # 12-provider universal client (18KB)
│ └── design-tokens.ts # Color + typography tokens
│
├── openclaw/ # OpenClaw Agent (CIK model)
│ ├── SOUL.md / IDENTITY.md / MEMORY.md
│ └── skills/ (3 skills)
│
├── graphrag/ # Python Backend
│ └── layers/
│ ├── graph_layer.py # TigerGraph schema + GSQL
│ ├── orchestration_layer.py # Dual pipeline + adaptive router
│ ├── llm_layer.py # LLM interactions
│ ├── evaluation_layer.py # RAGAS + F1/EM
│ └── universal_llm.py # LiteLLM 12-provider support
│
├── tests/test_core.py # 31 unit tests
├── Dockerfile # One-command deployment
└── README.md
📚 References
- GraphRAG — From Local to Global
- LightRAG — Simple and Fast (34K⭐)
- OpenClaw — Personal AI Agent
- HotpotQA — Multi-hop QA
- RAGAS — RAG Evaluation
- Youtu-GraphRAG — Schema-Bounded
TigerGraph · Anthropic · Ollama · Groq · LiteLLM · Next.js · Recharts