@vojtechMicrosoft Research dropped a paper on why question vector search hits a wall with RAG. The system pulls relevant chunks fine but fails when the answer lives in the whole corpus instead of one document. Asking about main themes exposes this gap because no single chunk holds the full picture. Ten authors at Microsoft Research propose a fix where an LLM reads the entire corpus to build an entity graph of facts and relationships rather than paragraphs. It clusters related entities and pre-writes a summary for each one before anyone asks. At query time you don't search. Each summary gives a partial answer then those merge into the final one. This uses map-reduce over structure instead of similarity search over text. On million-token datasets it beat conventional RAG by a wide margin. The evaluation measures comprehensiveness and diversity not accuracy using another LLM as the judge. Indexing requires paying upfront to read everything. If users ask questions about the corpus rather than questions answered by one document, no embedding model solves the problem. The work is titled From Local to Global: A GraphRAG Approach to Query-Focused Summarization.

@vojtechMicrosoft Research dropped a paper on why question vector search hits a wall with RAG. The system pulls relevant chunks fine but fails when the answer lives in the whole corpus instead of one document. Asking about main themes exposes this gap because no single chunk holds the full picture. Ten authors at Microsoft Research propose a fix where an LLM reads the entire corpus to build an entity graph of facts and relationships rather than paragraphs. It clusters related entities and pre-writes a summary for each one before anyone asks. At query time you don't search. Each summary gives a partial answer then those merge into the final one. This uses map-reduce over structure instead of similarity search over text. On million-token datasets it beat conventional RAG by a wide margin. The evaluation measures comprehensiveness and diversity not accuracy using another LLM as the judge. Indexing requires paying upfront to read everything. If users ask questions about the corpus rather than questions answered by one document, no embedding model solves the problem. The work is titled From Local to Global: A GraphRAG Approach to Query-Focused Summarization.