User proposes using rerankers to retrieve nodes in knowledge graph RAG

A user on the Hugging Face forums proposes a Retrieval-Augmented Generation (RAG) approach where documents are stored as nodes in a knowledge graph and retrieved via a reranker. The query itself becomes a new node that connects to the most similar nodes identified by the reranker.

• Documents or text sections are stored as nodes in a knowledge graph.
• A reranker retrieves documents based on a query, offering potentially higher precision than contextual RAG.
• The query becomes a new node establishing connections with the most similar nodes.
• This method addresses the limitation of LLMs used for node extraction, which may only generate coarse-grained labels or keywords.

The user notes that rerankers might be more precise than contextual RAG, while LLM-based node extraction can produce labels that are too coarse-grained.