PharmaIntel: GraphRAG Inference Benchmark

TigerGraph GraphRAG Hackathon 2026 – Round 1 Submission

Team: Expelliarmus (Solo) Tech Stack: TigerGraph Savanna, Groq Llama 3.3 70B, ChromaDB, Streamlit

Quick Link

• Blog Post: [https://dev.to/sachitha_srin_d4fba6/how-graphrag-cut-our-llm-token-costs-by-62-on-indian-pharma-data-43f9]
• Live Video: [https://www.loom.com/share/d64db4d2cdad448ab5b6a1a565362aad]

Architecture

Pipelines

1. LLM-Only: Groq + raw prompt
2. Basic RAG: Sentence-Transformers + ChromaDB + Groq
3. GraphRAG: TigerGraph Savanna knowledge graph + multi-hop traversal + Groq synthesis

Results

• Token Reduction: 62% vs Basic RAG, 79% vs LLM-Only
• Cost per query: $0.00048 (GraphRAG) vs $0.00126 (Basic RAG)
• Accuracy: 91% LLM-Judge pass rate, 0.72 BERTScore F1