eres45/savannaflow

SavannaFlow is a benchmarking platform that compares three different approaches to answering questions with AI: using a language model alone, using a standard method called Basic RAG, which stores text as vectors for similarity search, and GraphRAG, which stores information as a connected graph of relationships. The project was built for a hackathon using TigerGraph Savanna, a graph database, as its backbone. The core problem it addresses is what the README calls the Vector RAG Tax. When standard RAG systems search for an answer, they retrieve entire paragraphs of text, often over a thousand tokens, even when the actual answer is a single fact. SavannaFlow's GraphRAG approach instead traverses a graph of specific nodes and relationships, fetching only the precise data needed. For example, to find which contractor built a particular rocket engine, it follows a direct path through the graph rather than loading unrelated NASA history text. The platform uses NASA aerospace data as its test dataset. Users can submit a question and see all three approaches run side by side in real time, with metrics like token usage, cost per query, and accuracy displayed together. According to the benchmark results in the README, GraphRAG achieves roughly 3.5 times fewer tokens than Basic RAG while maintaining the same accuracy. The backend is written in Python using FastAPI, and the frontend dashboard uses Next.js. AI responses are generated through Groq, running an open-source large language model. The project also includes an automated accuracy scorer that acts as an AI judge to evaluate each answer, and a live demo dashboard is deployed for public testing.