Graph RAG Glossary
Abstract Concepts
An idea, quality, or notion that exists in thought or as an intellectual construct rather than as a physical or tangible object.
Unlike concrete concepts, which refer to things that can be perceived through the senses (such as “tree” or “car”), abstract concepts represent intangible entities like emotions, relationships, moral values, theories, and principles. Examples of abstract concepts include ideas like “freedom,” “justice,” “happiness,” and “love.” These concepts often require a deeper level of cognitive processing and are critical in understanding complex, philosophical, or theoretical ideas.
Examples of Abstract Concepts:
- Emotions and Feeling: happiness, sorrow, anger, pain
- Qualities: honesty, bravery, intelligence
- Theories: evolution, relativity
- Values: freedom, justice, equality
- States: peace, chaos, order
Even small language models are good finding concrete entities. However, there is additional processing time and cost associated with extracting abstract concepts from text and inserting them into a graph. Qualities may also be inferred from adjectives that describe a person such as "an honest person".
Community Detection
Community detection is the process of identifying groups of nodes within a network that are more densely connected to each other than to the rest of the network. This technique is used to reveal the modular structure of networks, uncovering clusters or communities that share common properties or functions.
For example, in social network analysis, community detection can help identify groups of users who frequently interact with each other, such as friend circles or interest groups.
Concept Network
A structured representation of related terms and concepts that captures the relationships between them, such as synonyms, alternate terms, abbreviations, broader terms, and narrower terms.
Concept networks serve as a semantic map that links various expressions of an idea or entity, enabling the system to recognize and interpret different linguistic variations of the same concept or to bind a concept with opposite terms.
Document Chunking
Document chunking is the process of dividing a large document into smaller, manageable sections or chunks. This technique is commonly used in natural language processing to facilitate tasks such as information retrieval, summarization, or content analysis by processing these smaller chunks individually.
For example, in a question-answering system, document chunking enables the model to focus on specific sections of a document, improving accuracy and efficiency in finding relevant answers.
Complex Sequential Question Answering (CSQA)
Complex Sequential Question Answering refers to a task in natural language processing that involves answering a series of interrelated questions where the context or the answers to previous questions impact subsequent questions. This task requires handling complex dependencies and reasoning chains, often relying on structured knowledge bases or graphs to provide accurate and context-aware responses. In GraphRAG, CSQA tasks benefit from the ability to retrieve and reason over graph-structured data that captures these intricate relationships(GraphRAG-Survey).#### Entity Extraction
Concrete Entities
Entities the correspond to physical items in the world. These are often referred to as POLE entities (People, Organizations, Locations and Events). These are often proper names and it is usually easy for even small language models to identify them.
Concrete entities are often contrasted with abstract [Abstract Concepts]
Entity Extraction
The process of identifying and categorizing key pieces of information (entities) from text, such as names of people, organizations, locations, events, and other concepts.
These entities are often represented by proper nouns or specific phrases that are crucial to understanding the text's meaning.
GNN-based Retriever
GNN-based retrievers use Graph Neural Networks to understand and process graph structures. They encode graph data, score different retrieval granularities based on similarity to the query, and guide the retrieval process within GraphRAG. These retrievers are particularly adept at handling complex graph data structures(GraphRAG-Survey).
Graph Neural Networks (GNN)
GNNs are a class of deep learning models designed to handle graph-structured data. They operate by aggregating information from a node's neighbors to generate node embeddings, which can then be used for various tasks, such as node classification or graph-based retrieval. In GraphRAG, GNNs are often employed to encode graph data and guide retrieval processes(GraphRAG-Survey).
Graph Retrieval-Augmented Generation (GraphRAG)
A methodology that enhances Retrieval-Augmented Generation (RAG) by leveraging graph data structures. GraphRAG retrieves elements such as nodes, triples, paths, or subgraphs from a pre-constructed graph database, allowing for more precise and context-aware generation by considering the interconnections between texts(GraphRAG-Survey).
Graph-Based Indexing
Graph-Based Indexing is a method for organizing and indexing graph data to optimize retrieval operations. This involves mapping node and edge properties, establishing pointers between connected nodes, and organizing the data to support efficient traversal and retrieval. It plays a crucial role in enhancing query efficiency in GraphRAG(GraphRAG-Survey).
Graph-Enhanced Generation
Abbreviation: G-Generation
The phase in GraphRAG where meaningful outputs or responses are synthesized based on the retrieved graph data. This stage involves taking the query, retrieved graph elements, and an optional prompt as inputs to generate a contextually relevant and accurate response(GraphRAG-Survey).
GraphRAG Step
The key steps in a GraphRAG process, as outlined in the Summary paper, are:
- Graph-Based Indexing (G-Indexing)
This initial phase involves constructing or identifying a graph database that aligns with the downstream tasks. The indexing process typically includes mapping node and edge properties, establishing pointers between connected nodes, and organizing the data to support fast traversal and retrieval operations. The quality of this step directly impacts the efficiency and effectiveness of the retrieval process.
- Graph-Guided Retrieval (G-Retrieval)
Following the indexing phase, the retrieval phase focuses on extracting relevant information from the graph database based on the user query. This involves selecting the most pertinent graph elements, such as entities, triplets, paths, or subgraphs, using retrieval algorithms. The goal is to find the graph elements that best match the query and support the next stage of generation.
- Graph-Enhanced Generation (G-Generation)
In the final phase, the retrieved graph elements are used to generate meaningful responses or outputs. This stage involves taking the query, the retrieved graph elements, and an optional prompt as inputs for a generation model, which produces the final response. The process is enhanced by the structural and relational knowledge embedded in the graph data.
These steps outline the comprehensive workflow of GraphRAG, emphasizing the importance of each phase in achieving accurate, contextually relevant responses based on graph-structured data(GraphRAG-Survey).
Hallucination
Hallucination in natural language processing refers to the generation of content by a model that is factually incorrect, nonsensical, or not supported by the input data. This occurs when the model produces information that appears coherent but is either fabricated or irrelevant to the context.
For example, a language model might hallucinate by generating a detailed but entirely fictitious historical event when asked about a specific time period.
Hybrid Granularities
Hybrid Granularities involve retrieving graph data at multiple levels of detail, such as nodes, triplets, paths, and subgraphs, to capture both fine-grained and broad contextual information. This approach enhances the relevance and comprehensiveness of the retrieved data, making it useful for complex queries within GraphRAG (GraphRAG-Survey).
ISO Definition
A term definition is considered to be consistent with ISO metadata registry guideline 11179 if it meets the following criteria:
- Precise
- Concise
- Distinct
- Non-circular
- Unencumbered with business rules
Knowledge Base Question Answering
Abbreviation: KBQA
A task in natural language processing focused on answering user queries based on external knowledge bases. KBQA methods are typically categorized into Information Retrieval (IR)-based methods, which retrieve information from knowledge graphs, and Semantic Parsing (SP)-based methods, which generate logical forms to query knowledge bases(GraphRAG-Survey).
Knowledge Merging
A post-retrieval enhancement strategy in GraphRAG that involves combining multiple pieces of retrieved information to provide a comprehensive view. This process assists in consolidating knowledge, improving the relevance and completeness of the final retrieved results(GraphRAG-Survey).
Knowledge Pruning
A technique used after initial retrieval to refine and improve the quality of the retrieved information by removing redundant or less relevant data. It aims to ensure that the final results are highly pertinent to the user's query, enhancing the relevance of the information presented(GraphRAG-Survey).
LM-based Retriever
LM-based retrievers leverage the natural language processing capabilities of language models to retrieve relevant graph data based on text queries. These models are particularly effective for interpreting and processing diverse queries in GraphRAG. They are used to generate reasoning paths and assist in retrieving the most relevant graph elements(GraphRAG-Survey).
Language Models
Abbreviation: LMs
Language models (LMs) are advanced AI models designed to understand and generate human language. They are mainly classified into two types: discriminative models, like BERT and RoBERTa, which focus on tasks such as text classification, and generative models, like GPT-3 and GPT-4, which are used for tasks like machine translation and text generation(GraphRAG-Survey).
Leiden Algorithm
The Leiden Algorithm is an iterative method used for community detection in large networks. It improves upon the Louvain algorithm by guaranteeing well-connected communities, reducing the likelihood of disconnected or fragmented clusters within the detected communities.
For example, the Leiden Algorithm is often applied in graph analysis to identify groups of nodes that are more densely connected internally, such as in social network analysis to find clusters of users with similar interests.
Leiden algorithm, are used to partition the graph into communities of nodes with stronger connections to one another than to other nodes in the graph. This helps in grouping closely related entities together
Non-parametric Retriever
Non-parametric retrievers are based on heuristic rules or traditional graph search algorithms. They do not rely on deep-learning models, making them efficient for retrieval tasks that require high speed. These methods typically involve pre-processing steps like entity linking and are used in GraphRAG for tasks like subgraph extraction(GraphRAG-Survey).
Paths
Paths represent sequences of relationships between entities within a graph. In GraphRAG, retrieving paths enhances contextual understanding and reasoning capabilities by capturing complex relationships and dependencies(GraphRAG-Survey).
Query Decomposition
A technique that breaks down the original user query into smaller, more specific sub-queries. Each sub-query focuses on a particular aspect of the original query, making it easier to retrieve pertinent information and reduce the complexity of language queries(GraphRAG-Survey).
Query Expansion
A pre-processing technique used in query enhancement to improve search results by supplementing or refining the original query with additional relevant terms or concepts. It aims to capture lexical variations and expand the information content of the query(GraphRAG-Survey).
Query-Focused Summarization
Abbreviation: QFS
A task in natural language processing where the goal is to generate a summary of a document or set of documents that is focused on a specific query. GraphRAG enhances QFS by retrieving relevant subgraphs that capture the broader context and interconnections within the graph structure(GraphRAG-Survey).
Retrieval-Augmented Generation
Abbreviation: RAG Retrieval-Augmented Generation is a methodology that integrates external knowledge retrieval with the generation process of language models. By dynamically querying a large text corpus, RAG enhances the contextual depth, factual accuracy, and relevance of the content generated by language models. It addresses challenges such as hallucinations, lack of domain-specific knowledge, and outdated information by incorporating relevant factual knowledge into the responses(GraphRAG-Survey)(GraphRAG-Survey).
Subgraphs
In GraphRAG, subgraphs refer to comprehensive subsets of a graph that encapsulate relational contexts within a larger structure. Retrieving subgraphs allows for capturing complex patterns, sequences, and dependencies, which are crucial for a deeper understanding of semantic connections in graph-based tasks(GraphRAG-Survey).
Synset
A group of words or phrases that have similar meaning.
For example, the words "baby", "infant" and "child" may be consider rough synonyms.
Text-Attributed Graphs
Abbreviation: TAGs
A universal format for graph data used in GraphRAG, where nodes and edges are associated with textual attributes. TAGs can represent knowledge graphs and other graph structures, enabling the retrieval of complex relational knowledge for use in generation tasks(GraphRAG-Survey).
Triplets
Triplets in graph databases consist of entities and their relationships, represented as subject-predicate-object tuples. This structured representation is used in GraphRAG to retrieve and understand relational data within a graph(GraphRAG-Survey).
Vector Indexing
Vector Indexing involves converting graph data into vector representations to facilitate fast retrieval. This is particularly useful for entity linking and rapid query processing, where vector search algorithms like Locality Sensitive Hashing (LSH) are applied. It enables efficient searches and helps maintain structural information during retrieval(GraphRAG-Survey).