memo: graphrag

[nnhuyhoang]

What's this PR do?

• memo for GraphRAG

[zlatanpham]

Suggested change

Naive RAG systems built with keyword or similarity search-based retrieval fail in complex queries that require reasoning. Suppose user asks a query: "Who directed the sci-fi movie where the lead actor was also in The Revenant?", a standard RAG system will retrieve documents about The Revenant, then find information about cast and crew members, but it will fail identiy the lead actor, Leo DiCaprio who starred in other movies but not existed in the query to retrieve document. However, taking advantage of knowledge graph, the ideally process will be: identify the lead actor of The Revenant, then find other movies he starred in, then find the sci-fi movie which meet ciriteria.

Naive RAG systems built with keyword or similarity search-based retrieval fail in complex queries that require reasoning. Suppose user asks a query: "Who directed the sci-fi movie where the lead actor was also in The Revenant?", a standard RAG system will retrieve documents about The Revenant, then find information about cast and crew members, but it will fail to identify the lead actor, Leo DiCaprio who starred in other movies but not existed in the query to retrieve document. However, taking advantage of knowledge graph, the ideally process will be: identify the lead actor of The Revenant, then find other movies he starred in, then find the sci-fi movie which meet criteria.