RAPTOR

RAPTOR – Recursive Abstraction from the Bottom Up:

1. document is broken down into leaf blocks;
2. semantically close blocks are clustered;
3. LLM generates a summary of each group as a parent node;
4. the process is repeated until the root.

The result is a knowledge tree that allows searching at different levels of abstraction: from a macro concept to details.

Suitable for questions of the type “gradually delve from concept to detail”.

Related: GraphRAG, RAG, Structured indexing