Quick breakdown of what GraphRAG is explained quickly, and simply.
What is GraphRAG
GraphRAG (Graph-based Retrieval-Augmented Generation) is a technique used to enhance LLM user-query responses by utilizing data from external knowledge graphs in addition to the trained dataset to generate more accurate and relevant responses.
GraphRAGs help in solving LLMs most common limitation being that the trained data might be outdated and may return responses that are inaccurate. GraphRAG is an enhancement to RAG (Retrieval-Augmented Generation) as it queries a knowledge graph instead of a vector text database which is then fed into the LLM. Using a graph allows for better querying for relationships between data rather than for just the data itself.
GraphRAG vs Alternative Approaches
The approach of technology depends on what the customer aims to accomplish with the integration. If the customer’s goal is to have a more detailed and semantic-based search and find code within a large codebase, then a semantic-based search might be more relevant.
In contrast, if the customer wants to understand the dependencies of a piece of code, a graph-based approach where relationships are mapped between functions may be more relevant.
The customer might also want to train their own model if their codebase is large and is a part of a legacy system where core functions are not updated as frequently (for example a banking system or airline guidance-systems).
The table below summarizes findings between other approaches that may be taken depending on the use-case.
| Approach | Strength | Weaknesses |
|---|---|---|
| GraphRAG | Understands relationships between different entities and can connect them together for more complex, conceptual-based queries | Might take longer to query since the context window is very large. Might struggle more on unstructured data since relationships cannot be mapped as explicitly. |
| Vector-based Text Search | Better at handling text-based queries where there exist different ways to say the same thing. Can infer “meaning” of a query better. Would be more useful to search for text | Is not able to efficiently determine relationships between entities. Only look for content that is “similar”. Better at querying unstructured data. |
| Prompting | Most easily accessible by users | Does not contain the whole context of the query. Very cumbersome to add a plethora of background information for a query. |