GraphRAG

GraphRAG is an advanced RAG pattern that combines an ingestion process for extracting entities and relationships from unstructured text and further using graph algorithms for enrichment and summarization. The retrieval step then uses the knowledge graph in combination with vector search to navigate to more relevant information than just the initial text chunks.

Traditional LLMs excel in conversational behavior and generative language skills but struggle with recent, private, or low-represented information, often leading to hallucinations, due to the focused training on helpfulness instead of factuality.

RAG architectures mitigate this by grounding LLM responses in trusted, current, and private information, improving factual correctness.

While LLMs show impressive results in tasks that require language skills, they are not trustworthy when it comes to answering factual questions even more so in business critical applications that rely on the correct and secure usage of private information.

They need a powerful companion data source that can capture real-world information at high fidelity and make the context-relevant direct and indirect facts available quickly and in a transparent manner. Knowledge Graphs managed in graph databases present that "memory" (left-brain) companion to the language (right-brain) models capabilities.

GraphRAG covers a set of advanced techniques for GenAI applications that improve quality and explainability of answers and task completions. GraphRAG applies to data preparation and retrieval (the R in RAG) that involve graph (connected) data structures (and graph databases).

Basic RAG architectures rely on vector retrieval of text chunks based on the vector embedding of the user’s question. Those retrieved chunks are then presented with the question to an LLM to generate a human answer. In most cases the quality and explainability of the basic approach is not sufficient so that more advanced RAG patterns are required.

In RAG architectures there are different patterns that are applied at different phases of the process to improve quality of the answer.

Some are listed below

Many advanced RAG patterns rely on connecting pieces of information to each other and using those references during retrieval for fetching additional related information. In a non-graph system that means both storing a lot of extra lookup ids and writing and maintaining the large amount of extra code to resolve those during retrieval.

In a graph system it is enough to add relationships between pieces of information and following the relationships directly or indirectly during retrieval as part of a query and using them during extraction for additional clustering and relating information. So this is the most natural operation in a graph while it is a bolted on, high effort activity in other systems.

There are a number of different techniques involved that all rely on linking relevant pieces of information in a way that they capture the real-world relationships to support answering questions by navigating this graph of knowledge.

To reach the trust level required for critical applications besides quality explainability and transparency is key - with graph data structures and graph queries during execution, the "black box" vector only search is replaced by a detailed trace of querying and a detailed set of results that go into answering the question. The facts use for answers can also be linked to the answer and question in the graph model to be used for auditing and future improvements (e.g. re-ranking based on user feedback for this class of questions).

For answering involved questions, vector and structural graph search are combined to capture relevant clusters of information so that the LLM has all the facts (and their origins) to answer the questions. Additionally at every stage this architecture allows digging deeper and expanding the search into adjacent areas.

GraphRAG is based on structuring information into relevant pieces and connecting them, this allows for making these specific elements accessible and addressable and make them part of a navigation across a network of information

While documents mostly represent a vertical (and sometimes temporal) structure of a flow of segments of information, GraphRAG with entity extraction and clustering captures the horizontal re-occurrence and enhancement of topics across documents.

Text attributes of structured data or unstructured documents are structured into chunks (can be document hierarchies). This is also referred to as the lexical graph

Entities of different types and their typed relationships are extracted from text chunks with a given graph schema (or automatic)

This is also referred to as the domain graph or entity graph

As extracted entities form clusters of information in the graph these clusters can be determined with graph algorithms (WCC, Louvain, Leiden) and used to aggregated larger set of entities into "topic" clusters that can both be named but also summarize entities and their relationships

Optionally other information like facts or claims on entities and relationships can be extracted from the text and stored in the graph

The graph can be enriched from other enterprise or public sources

This retrieval pattern combines scored vector search with structural graph search, with vector search text chunks are identified that are relevant to the question, then entities related to these chunks are extracted and from these entities relationships are followed further out (multi-hop) to find farther away, but related information

Here entities are retrieved by vector and/or full-text search based on the user’s question and then related information (text chunks, cluster summaries, claims, facts, …) is retrieved

This is used for answering global questions \- it uses the clustered topic summaries to capture the general themes of the dataset and then navigates from there to other relevant entities.

Benefits of using GraphRAG for better summarization and retrieval of information

Democratization of of access to information in a structured, trustworthy and explainable way, esp. useful for pre-existing knowledge graphs.