liveBook · Manning

G-indexing

As explained in Retrieval Augmented Generation, The Foundational Ideas

G-Indexing extracts a graph, organizes it into hierarchical communities, and precomputes summaries for later retrieval.

A question about an incident may depend on connections scattered across documents: who was on call, which team owned a service, and which incident affected it. Ordinary RAG indexes passages for similarity search, but the retrieved passages arrive independently. G-Indexing prepares a graph of entities and relationships so later retrieval can follow connections between them. It is the offline construction stage, before query-time G-Retrieval and G-Generation. section 12.1.1

The shared construction pattern is:

  • Split source documents into manageable chunks, then use an LLM to extract entities and relationships, often represented as subject–predicate–object triples. section 12.1.1
  • Build a graph from the extracted structure. Keep links to source chunks so graph elements can be traced back to passages. section 12.1.1 section 12.2.1

GraphRAG makes this index rich enough for broad summarization. For each chunk, its LLM extracts typed entities with descriptions and relationships with descriptions and numeric strengths. A bounded second-pass “gleaning” loop asks for missed elements. Repeated mentions are combined into node and relationship descriptions. The graph is then partitioned into communities at multiple levels, with a precomputed summary for each community. Query-time retrieval can use this hierarchy rather than asking the LLM to reconstruct an overview from disconnected chunks. section 12.2.1

From an on-call passage to graph evidence

A passage saying that Anna Müller owns AuthService and was on call during INC-2347 can yield a person node, a service node, an incident node, and relationships recording ownership and on-call responsibility. Each extraction remains tied to its source chunk; mentions from other passages can contribute to the same node’s description. section 12.2.1

Not every G-Indexing design builds a summary pyramid. HippoRAG extracts triples into a knowledge graph, links noun-phrase nodes to passages, and adds dense-encoder synonymy edges for variant names. Its nodes remain bare phrases rather than detailed profiles. That makes indexing substantially cheaper than GraphRAG’s description-and-summary pipeline, but leaves HippoRAG without summaries for questions such as “What are the main themes?” Choose the index according to the retrieval job: connected paths for multi-hop questions, or precomputed community views for corpus-level synthesis. section 12.2.1 section 12.3.2

How Retrieval Augmented Generation, The Foundational Ideas covers it

How deep it goes
Works through extraction, provenance, aggregation, and community summaries; contrasts that pipeline with a lighter graph index.
Examples
Uses an incident-and-service passage to illustrate typed entities, relationships, strengths, and links back to source chunks.
What it assumes
Builds on the chunk-and-embed RAG workflow while defining graph extraction and construction as separate indexing steps.
What you can do after
Evaluate/understand whether a summary-rich GraphRAG index or a cheaper associative HippoRAG index fits the intended questions.
How it connects
Connects knowledge graphs and entity extraction to G-Retrieval, community summarization, synonymy edges, and multi-hop retrieval.

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.