Introducing KET-RAG: A Cost-Efficient Indexing Framework Built on Microsoft GraphRAG (Similar to Lazy GraphRAG)

[jeremyzhangsq]

🚀 Introducing KET-RAG: A Cost-Efficient Multi-Granular Indexing Framework for Graph-RAG 🚀

Large-scale Graph-RAG indexing is expensive—so much so that Microsoft’s official repo warns users to "start small" to avoid unexpected costs. But just how expensive is it? In the legal industry, processing a single case (usually containing 5GB of text information) using GPT-4o-mini could cost $33K—a major bottleneck for real-world applications.

🔍 What’s the solution?

We propose KET-RAG, a novel multi-granular indexing framework that significantly reduces costs while maintaining (or even improving) retrieval quality. Our approach:

✅ Identify core text chunks using PageRank on an intermediate KNN graph
✅ Generate a knowledge graph skeleton from these chunks
✅ Construct a lightweight text-word bipartite graph using only word & sentence tokenization—no expensive embedding models needed!

⚡ Key Results

📌 Comparable or superior retrieval quality to Microsoft’s Graph-RAG
📌 An order of magnitude lower indexing costs

🔎 What’s next?

Our next steps include handling global search and tackling scalability & industrial deployment challenges. We welcome discussions and collaborations on pushing KET-RAG further!

🔗 Check out the full work:

📜 Technical Report

💻 Code & Implementation

Would love to hear your thoughts—how do you see this impacting real-world RAG applications? Let’s discuss! 💬 #AI #RAG #GraphRAG #KnowledgeGraphs #LLM #KETRAG

[errogaht]

interesting! does that means we don't need to wait lazy graph rag?

[jeremyzhangsq]

you can have a try:D
Please let me know if you have further question.