GraphRAG, the better alternative to RAG for AI chatbots and response engines, is now freely accessible on GitHub.
Microsoft is releasing GraphRAG, a new technique that surpasses regular Retrieval-Augmented Generation (RAG) by significant margins and allows chatbots and response engines to connect the dots across a whole dataset.
What Distinctions Exist Between GraphRAG and RAG?
Using a technique called RAG (Retrieval-Augmented Generation), an LLM may access a database, such as a search index, and utilize it as the foundation for a response. It can serve as a link between a traditional search engine index and a sizable language model.
The advantage of RAG is its ability to provide answers based on reliable and accurate data. Additionally, RAG makes it possible for chatbots using generative AI to leverage current data to respond to inquiries regarding subjects outside of the LLM’s scope of training. Perplexity and other AI search engines employ this strategy.
RAG’s use of embeddings contributes to its advantages. Embeddings represent the semantic links between words, phrases, and texts. The retrieval portion of RAG may match text in a database (such as a search index) with a search query.
However, the drawback of embedding is that it restricts the RAG to local text matching (instead of a broad reach over the data).
GraphRAG’s unique feature lets an LLM respond to inquiries using the entire dataset.
GraphRAG converts the indexed documents—also referred to as unstructured data—into a knowledge graph. Web pages are a clear example of unstructured data. In other words, while GraphRAG generates a knowledge graph, it does so by producing a more machine-understandable “structured” representation of the interactions between different “entities” (such as individuals, locations, ideas, and objects).
Using more specific subjects (low level) and broad themes (high level), GraphRAG generates what Microsoft refers to as “communities.” Subsequently, an “LLM produces a hierarchical summary of the data,” which is then used to address inquiries from a summary of these communities. The innovation allows a chatbot to respond to queries less reliant on embeddings and more on knowledge (the summaries).
Examples of RAG Versus GraphRAG
The original GraphRAG research report demonstrated the advantage of the GraphRAG technique in answering
queries without precise match data in the indexed texts. The example uses a small dataset of June 2023 Russian and Ukrainian news stories (translated into English).
GraphRAG is Now Open to the Public
Microsoft said that anybody can use GraphRAG in the public domain.
Microsoft published GraphRAG to provide public accessibility to the solutions built upon it and to invite suggestions for enhancements.
