Graph-RAG systems answer questions by retrieving from a knowledge graph extracted from a document corpus — but building that graph is expensive, because every text chunk must pass through an LLM to extract entities and relations. For large corpora, indexing cost alone can dominate the bill.

The idea

KET-RAG (Knowledge-Enhanced Text RAG) observes that not all chunks deserve the same indexing budget. It builds a multi-granular index: a full knowledge-graph skeleton over the most influential text chunks (selected via a PageRank-style importance measure), plus a lightweight keyword–chunk bipartite graph over the rest. Retrieval then combines both layers — structured traversal where the skeleton exists, and cheap keyword matching elsewhere.

Why it matters

This design keeps the retrieval quality benefits of Graph-RAG at a fraction of the indexing cost, offering a tunable dial between the cheap-but-flat classic RAG and the rich-but-costly full Graph-RAG. On standard benchmarks, KET-RAG matches or beats full Graph-RAG indexing quality while cutting LLM indexing cost by an order of magnitude.