liveBook · Manning

Intrinsic structure

As explained in Retrieval Augmented Generation, The Foundational Ideas

Intrinsic structure is the preexisting hierarchical organization explicitly provided by a document, such as chapters, sections, or numbered filing sections.

Intrinsic structure is the hierarchy already supplied by a document, such as chapters, sections, subsections, numbered filing sections, and page ranges. Structure-aware retrieval uses this organization instead of inducing a hierarchy through unsupervised clustering. section 10.7.3

Choosing document structure over clustering

For a textbook or SEC filing with clean structural metadata, map retrieval to its table of contents rather than constructing a synthetic hierarchy. This can avoid cluster incoherence and is presented as more production-relevant than unsupervised RAPTOR-style organization. section 10.7.3

How Retrieval Augmented Generation, The Foundational Ideas covers it

How deep it goes
Defines the design choice and contrasts it with recursive clustering, but does not detail implementation algorithms.
Examples
Uses textbooks, SEC filings, tables of contents, subsections, and page ranges as structural examples.
What it assumes
Assumes access to clean structural metadata and distinguishes it from hierarchies induced from unstructured text.
What you can do after
Evaluate retrieval architectures by preferring document-provided hierarchy when structural metadata is clean and reliable.
How it connects
Connects hierarchical retrieval with table of contents, unsupervised clustering, embeddings, entity trees, and document metadata.

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.