Graph-RAG Framework Launches for Complex Queries

## Graph-RAG: Under the Hood of the New Open-Source Framework The new Graph-RAG framework, introduced by a team at Microsoft Research led by Jonathan Larson, fundamentally shifts how retrieval systems handle complex queries by creating a knowledge graph from unstructured text. This graph-based data structure allows for multi-hop reasoning, connecting disparate pieces of information in a way that traditional vector search struggles with, leading to more comprehensive and contextually-aware answers. Unlike standard RAG that retrieves isolated text chunks based on semantic similarity, Graph-RAG first uses an LLM to extract entities and their relationships from documents, building a structured knowledge graph. When a query comes in, the system traverses this graph to find interconnected data, providing the LLM with a richer, more structured context to generate a response. This method has shown significant performance improvements in benchmarks, with some studies reporting a 3x increase in accuracy on complex, multi-hop questions. From a cost perspective, implementing the open-source Graph-RAG framework involves a trade-off. The initial construction of the knowledge graph can be computationally expensive, with some estimates suggesting that about 75% of the total indexing cost comes from the LLM calls needed for entity and relationship extraction. However, once the graph is built, per-query costs can be lower and more efficient than traditional RAG, which may require processing large volumes of text for each query. The total cost of ownership for a "build" approach using the open-source framework will also need to factor in engineering effort for implementation, hosting of the graph database, and ongoing maintenance. The launch of an open-source Graph-RAG framework introduces a new dynamic into the enterprise search market. Competitors like Glean also leverage a knowledge graph, but with an added "personal graph" that tailors results based on an individual's role and interactions within the company. Glean's pricing is reported to start at over $50 per user per month, with the total cost of ownership potentially being two to six times the initial license fee when factoring in AI add-ons and infrastructure overhead. Other players in the enterprise search space have different architectural approaches. Hebbia, which targets the finance and legal sectors, utilizes a multi-agent AI system. This approach breaks down complex user questions into smaller sub-tasks that are then assigned to different agents for resolution. Their pricing is not public but is reportedly high, involving a lengthy sales process and long-term contracts. Cohere, another major competitor, focuses on providing a powerful "Rerank" model as part of its offering. This model takes the initial search results from a retrieval system, which could be vector-based or otherwise, and re-orders them to improve relevance before feeding them to a generative model. Cohere offers a more transparent, token-based pricing model for its various models, allowing for more predictable costs based on usage. The decision for an enterprise to build on an open-source framework like Graph-RAG versus buying a solution from a vendor like Glean, Hebbia, or Cohere will depend on several factors. A "build" approach offers greater customizability and control, but requires a significant upfront investment in engineering resources and infrastructure. A "buy" approach provides a managed solution with support, but comes with licensing costs and potential vendor lock-in. The availability of a robust open-source option now gives engineering teams a credible path to developing sophisticated, in-house enterprise search capabilities.