Knowledge graphs for AI

A graph makes entities and their connections explicit. In RDF, each statement is a subject, predicate, and object; a set of those triples forms a graph. The predicate expresses the relationship between the subject and object, while the object can be another resource or a value such as a literal. This is one formal graph model, not a requirement that every AI knowledge graph use RDF.

When modeling a retrieval problem, decide what counts as an entity, which relationships matter to the questions, and how each extracted statement points back to its source. GraphRAG’s dataflow documentation describes text units as chunks used for graph extraction and says extracted knowledge items reference them for provenance. A graph with unclear entity identity or unsupported relationships can make retrieval harder to inspect rather than easier.

• GraphRAG indexing overview

  Shows an AI retrieval pipeline that extracts entities, relationships, and claims from text.

Knowledge graph

A structured representation of entities and their relationships.

GraphRAG

A family of RAG approaches using graphs to organize or retrieve information.

Graph retrieval is worth evaluating when a question depends on connections among entities or evidence spread across a collection. GraphRAG documentation describes local search as combining graph data with raw-document text chunks for questions about specific entities. It describes global search as searching community reports for questions about the dataset as a whole, and includes basic vector RAG as a comparison option. The Microsoft Research paper presents GraphRAG as an approach to global sensemaking questions over a text corpus.

Use a question set that reflects the actual task, then compare graph retrieval with the baseline using the same questions and answer conditions. Check whether the retrieved passages support the answer, especially for multi-hop questions. A graph index adds extraction and organization steps; Microsoft’s documentation notes a cost and noisiness tradeoff between its indexing methods. Treat results as evidence for this corpus and task, not as a universal ranking.

• From Local to Global: A Graph RAG Approach to Query-Focused Summarization

  Presents a Graph RAG approach for questions that require summarizing a corpus.

• Indexing Methods

  Describes indexing approaches and their tradeoffs in graph detail and noisiness.

What Is Graphrag

Provenance records information about how data or another result was produced. The W3C PROV model organizes provenance around entities, activities, and agents, with relations for use, generation, derivation, and attribution. Those links can describe which source material an extraction activity used, what statement it generated, and who or what was responsible for the activity.

For retrieval, retain the source document and passage associated with an extracted entity, relationship, or claim. That allows a reviewer to inspect the evidence and the transformation path. A provenance record documents lineage; it does not certify that a source or extracted claim is true. Check the original passage and its context before relying on a retrieved statement.

• PROV Model Primer

  Introduces provenance concepts and how provenance assertions can be represented.