liveBook · Manning

HiRAG

As explained in Retrieval Augmented Generation, The Foundational Ideas

HiRAG is a hierarchical retrieval method that applies Gaussian mixture clustering to an entity graph rather than to text chunks.

HiRAG is a hierarchical retrieval method that applies Gaussian mixture clustering to an entity graph rather than to unstructured text chunks. It therefore differs from RAPTOR's recursive text clustering, while its name should not be confused with BookRAG's table-of-contents-based hierarchy. section 10.7.3

The architectural distinction is:

  • RAPTOR induces hierarchy by recursively clustering text representations. section 10.7.3
  • HiRAG applies the same Gaussian mixture clustering idea to an entity graph. section 10.7.3
  • BookRAG instead maps retrieval onto a document's inherent table of contents, such as chapters and sections. section 10.7.3

Choosing the hierarchy

For a textbook or SEC filing with reliable chapter or section metadata, an intrinsic hierarchy can outperform unsupervised clustering. HiRAG remains the entity-graph alternative; it does not directly exploit the table of contents. section 10.7.3

How Retrieval Augmented Generation, The Foundational Ideas covers it

How deep it goes
Defines the clustering substrate and contrasts HiRAG with RAPTOR and BookRAG, but gives no implementation procedure or evaluation results.
Examples
Contrasts textbook and SEC-filing hierarchies with entity-graph and table-of-contents indexing.
What you can do after
Evaluate whether entity-graph clustering or intrinsic document structure better fits a corpus than recursive text clustering.
How it connects
Connects HiRAG to RAPTOR, BookRAG, Gaussian mixture clustering, entity graphs, and intrinsic document structure.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Retrieval Augmented Generation, The Foundational Ideas. The section links point to the parts of the book it draws on.