How does Graph RAG combine structured and unstructured data?

Graph RAG (Retrieval-Augmented Generation) is an advanced technique that leverages knowledge graphs to integrate and combine both structured and unstructured data, providing a richer context for Large Language Models (LLMs) and improving the accuracy and relevance of generated responses.

The Challenge with Unstructured Data

Unstructured data, such as text documents, articles, emails, or web pages, contains a wealth of information but lacks a predefined data model. While LLMs excel at processing and understanding natural language, retrieving precise facts and relationships from vast unstructured corpuses can be challenging, often leading to hallucinations or incomplete answers, especially when specific entities and their connections are crucial.

The Value of Structured Data

Structured data, typically found in databases, spreadsheets, or APIs, is organized in a tabular format with defined relationships between entities. This data is precise, queryable, and consistent, making it excellent for fact-checking and specific data retrieval. However, structured data often lacks the nuanced context and explanatory power present in natural language.

How Graph RAG Combines Data Types

Graph RAG's core mechanism for combining data types lies in its use of a knowledge graph. The knowledge graph acts as a central, unified representation that can accommodate and link information from both structured and unstructured sources.

Processing Unstructured Data for the Graph

For unstructured data, advanced Natural Language Processing (NLP) techniques are employed. This involves:

• Information Extraction: Identifying key entities (e.g., people, organizations, locations, events) and their relationships within the text.
• Entity Linking/Resolution: Mapping extracted entities to existing nodes in the knowledge graph or creating new ones, ensuring consistency.
• Fact Triples: Converting natural language sentences into structured triples (subject-predicate-object) that can be represented as nodes and edges in the graph.
• Contextual Embeddings: Storing vector embeddings of text chunks alongside graph nodes, allowing for semantic similarity searches.

Crucially, the original unstructured source documents or specific text spans are often linked to the corresponding nodes and edges in the graph. This allows for direct retrieval of the original textual context when needed.

Integrating Structured Data Directly

Existing structured data from databases, APIs, or CSV files can be directly ingested into the knowledge graph. This data is typically already in a format (e.g., tables, objects with properties) that maps naturally to nodes (entities) and edges (relationships) within the graph schema. For instance, a database table of 'Products' with 'Price' and 'Manufacturer' fields can become 'Product' nodes connected via 'has_price' and 'manufactured_by' edges to 'Price' and 'Manufacturer' nodes, respectively.

The Unified Retrieval Process

When a user query is received, Graph RAG performs a two-pronged retrieval:

• Graph Traversal/Query: The query is used to traverse the knowledge graph, identifying relevant entities, relationships, and attributes. This provides highly precise, structured context, answering questions like 'Who founded company X?' or 'What is the relationship between Y and Z?'. This part leverages the strength of the structured data.
• Semantic Search/Unstructured Retrieval: Based on the graph context and potentially the initial query, relevant chunks of unstructured text linked to the identified graph entities are retrieved. This provides the descriptive and nuanced textual explanations that structured data alone might lack.

Both the precise, structured facts from the graph and the semantically relevant snippets from the unstructured documents are then combined and augmented into the LLM's prompt. This comprehensive context allows the LLM to generate more accurate, grounded, and coherent responses, bridging the gap between specific data points and natural language explanations.