Lexical Graph with Extracted Entities

Context

The biggest problem with the vector search approach as in the e.g. Basic Retrievers or Parent-Child Retrievers is finding all relevant context that is necessary to answer a question. The context can be spread across many chunks not being found by the search. Relating the real-world entities from the chunks to each other and retrieving these relationships together with a vector search provides additional context about these entities that the chunks deal with.

Graph Pattern

Elements

Nodes

Document nodes contain the document name and its source. They may contain additional metadata. Chunk nodes contain the human readable text of a chunk and its vector embedding. They may contain additional metadata. Entity nodes contain the name of the entity. Additionally they might contain a description of the entity and a vector embedding (of name and description). They may contain additional metadata. Entity nodes can have additional labels based on the extraction prompt.

Relationships

The PART_OF relationships do not require additional properties. However, they may contain additional metadata.

The HAS_ENTITY relationships do not require additional properties. However, they may contain additional metadata.

The RELATES_TO relationships are examplary here. They can have any type based on the extraction prompt. They may also have a description. They may contain additional metadata.

Description

Use an LLM to execute entity and relationship extraction on the chunks. Import the retrieved triples into the Graph.

GraphRAG Pattern

A Lexical Graph with Extracted Entities is used in the Graph-Enhanced Vector Search.