TigerGraph GraphRAG Inference Hackathon - Round 1 Submission
A production-grade, multi-pipeline benchmarking platform proving GraphRAG's superiority in token efficiency, latency, and accuracy over traditional vector-based RAG.
GraphRAG is fundamentally superior to vector-based Semantic RAG for complex, multi-hop reasoning tasks, especially in highly relational domains like Biomedicine.
To prove this definitively, we built an asynchronous three-pipeline benchmarking system. By querying a shared biomedical corpus, we fanned out exact queries to:
- Pipeline 1: LLM-Only Baseline (Zero-Shot via Groq)
- Pipeline 2: Basic RAG (Qdrant Vector DB + LlamaIndex + Groq)
- Pipeline 3: GraphRAG (TigerGraph Savanna SupportAI)
The Result: TigerGraph GraphRAG achieved a ~78% reduction in token usage and bypassed API rate limits entirely by utilizing structured graph traversal instead of stuffing massive, noisy text chunks into an LLM context window.
Our architecture adheres to Hexagonal Architecture principles, strictly isolating HTTP delivery, orchestration logic, and downstream pipeline execution. All services are fully containerized.
graph TD
%% Styling
classDef user fill:#e1f5fe,stroke:#0288d1,stroke-width:2px,color:#000
classDef orchestrator fill:#fff3e0,stroke:#f57c00,stroke-width:2px,color:#000
classDef pipeline fill:#e8f5e9,stroke:#388e3c,stroke-width:2px,color:#000
classDef db fill:#fce4ec,stroke:#c2185b,stroke-width:2px,color:#000
classDef llm fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px,color:#000
classDef dashboard fill:#e0f2f1,stroke:#00796b,stroke-width:2px,color:#000
%% Core Flow
User[User / Judge]:::user --> |"UI Interaction"| Dash[Glassmorphic Dashboard]:::dashboard
Dash --> |POST /v1/ask| Orch[FastAPI Orchestrator]:::orchestrator
%% Parallel Fan-Out
Orch --> |Asyncio Gather| P1[Pipeline 1: LLM-Only]:::pipeline
Orch --> |Asyncio Gather| P2[Pipeline 2: Basic RAG]:::pipeline
Orch --> |Asyncio Gather| P3[Pipeline 3: GraphRAG]:::pipeline
%% Downstream Services
P1 --> |Direct Prompt| Groq1[Groq API]:::llm
P2 --> |1. Semantic Search| Qdrant[(Qdrant Vector DB)]:::db
P2 --> |2. Context Prompt| Groq2[Groq API]:::llm
P3 --> |1. Entity Extraction| TG[(TigerGraph SupportAI)]:::db
TG --> |2. Graph Traversal| GraphLogic{GSQL Graph Logic}
GraphLogic --> |3. Synthesize| LLM3[Internal LLM]:::llm
We ran highly complex biomedical queries (e.g., "How does Alzheimer's disease affect acetylcholine neurotransmission?") across the suite.
| Metric | Pipeline 1 (LLM-Only) | Pipeline 2 (Basic RAG) | Pipeline 3 (GraphRAG) |
|---|---|---|---|
| Token Usage | ~316 tokens | ~1,340 tokens (Bloated) | ~287 tokens (Efficient) |
| Cost (USD) | ~$0.000125 | ~$0.000246 | $0.000000 (Managed) |
| Accuracy (Multi-hop) | Hallucination Risk | Frequent "Not enough info" | Precise & Relational |
Conclusion: Basic RAG relies on text chunks ranking highly in vector space. If a gene name isn't semantically close to a disease name, Qdrant fails to retrieve the critical chunk. GraphRAG circumvents semantic limitations entirely by walking deterministic graph edges (e.g.,
Disease→ASSOCIATED_WITH→Gene), returning 100% accurate context with a massive token discount.
This project is built for immediate, reproducible deployment using Docker Compose.
- Docker Desktop & Docker Compose
- TigerGraph Savanna Workspace Credentials
- Groq API Key
- Qdrant Cloud Cluster Key
Clone the repository and set up your environment variables:
git clone https://github.com/your-username/Graph-RAG.git
cd Graph-RAG
cp .env.example .envEdit .env to insert your active API keys.
We use Docker Compose to spin up the entire orchestrated fleet:
cd infra
docker compose up -d --buildFor production deployment, the architecture naturally fits a microservices paradigm:
- Frontend Dashboard: Deployable as a static site via Vercel, Netlify, or AWS Amplify.
- FastAPI Backend Services: (Orchestrator, LLM-Only, Basic RAG) can be deployed via Railway, Render, or an AWS EC2 / DigitalOcean Droplet utilizing the provided
docker-compose.yml. - Database Layer: TigerGraph is natively hosted on tgcloud.io (Savanna), and Qdrant is hosted on Qdrant Cloud.
Navigate to http://localhost:3000 (or your mapped dashboard port) to view the live side-by-side benchmarking UI.
Graph-RAG/
├── dashboard/ # Premium glassmorphic UI (Vanilla JS + CSS)
├── infra/ # Docker Compose, networking, Nginx configs
├── scripts/ # Automated data ingestion & test scripts
└── services/
├── orchestrator/ # FastAPI async fan-out & cost calculator
├── llm-only/ # Pipeline 1: Zero-shot baseline
└── basic-rag/ # Pipeline 2: Qdrant vector retrieval