SAGE framework boosts RAG recall by 5.7-8.5 points for multi-modal data
New AI retrieval method constructs chunk-level graphs to find evidence chains across text, tables, and graphs.
Deep Dive
Researchers Prasham Titiya, Rohit Khoja, Tomer Wolfson, Vivek Gupta, and Dan Roth developed SAGE (Structure Aware Graph Expansion), a framework for retrieval-augmented generation (RAG) over heterogeneous data. It builds a chunk-level graph offline and expands from seed chunks at query time. On benchmarks OTT-QA and STaRK, SAGE improved retrieval recall by 5.7 and 8.5 points over standard flat similarity search methods.
Why It Matters
Enables more accurate AI answers by connecting evidence across different data formats like documents, spreadsheets, and knowledge graphs.