Graphwise aims to boost AI accuracy with GraphRAG launch
With standard RAG pipelines proving unreliable, the vendor's new feature uses knowledge graphs to add needed context to the data retrieval process that fuels AI outputs.
With retrieval-augmented generation pipelines struggling to deliver the relevant data that agents and other AI applications need to deliver trustworthy outputs, Graphwise launched GraphRAG to provide customers with an alternative designed to enable more successful AI development.
Retrieval-augmented generation (RAG) is a framework for connecting applications such as agents and chatbots with data sources. However, with most AI initiatives failing to make it past the pilot stage and into production, standard RAG pipelines haven't proven good enough on their own to enable enterprises to deliver usable, trustworthy AI tools.
In January, Databricks launched Instructed Retriever, an alternative to RAG that adds more context to data retrieval such as user instructions and previous examples. Traditional RAG systems only use a user's query.
Graphwise's GraphRAG, which was released on Feb. 16, unites agents and other applications with a knowledge graph that acts as a semantic layer and is similarly aimed at improving on standard RAG pipelines.
Alan Morrison, an independent analyst, noted that while knowledge graphs can be traced back to the 1960s, they are taking on greater importance because agents need the context knowledge graphs provide to perform to enterprise standards. As a result, GraphRAG is a significant addition for Graphwise users.
"Graphwise can bring all enterprise data, content and knowledge together using standard-based graph description logic that's been around for decades, but only now becoming indispensable because the agent paradigm is here, and agents desperately need reliable context," he said. "With GraphRAG, tapping the power of that contextualized data becomes simpler."
Stephen Catanzano, an analyst at Omdia, a division of Informa TechTarget, likewise noted that GraphRAG is valuable for Graphwise users given that it combines graph technology and RAG.
"Bringing them together is powerful," he said. "GraphRAG is a significant addition for Graphwise customers as it enables them to leverage knowledge graphs as a semantic backbone, ensuring AI responses are grounded in verifiable enterprise facts and complex relationships. This is something standard RAG systems struggle to achieve."
With a North American headquarters in New York City and a European headquarters in Sofia, Bulgaria, Graphwise is a graph technology vendor formed in 2024 when Ontotext merged with Semantic Web Company. Competitors include specialists such as Neo4j and TigerGraph as well as broader-based database providers featuring graph database capabilities including AWS, Google Cloud and Microsoft.
Improving AI development
Problems related to data aren't the sole reason most AI initiatives fail before making it into production. Unrealistic expectations, lack of a clear business case and difficulties integrating applications into real-world workflows are among the other reasons an estimated 80% of all AI projects fail.
GraphRAG is a significant addition for Graphwise customers as it enables them to leverage knowledge graphs as a semantic backbone, ensuring AI responses are grounded in verifiable enterprise facts and complex relationships. This is something standard RAG systems struggle to achieve.
Stephen CatanzanoAnalyst, Omdia
But issues with data -- including the lack of sufficient high-quality, relevant data -- are among the main ones.
Graphwise's new feature targets the discovery of relevant data. Agents and other applications are built for specific tasks. For example, many enterprises are building agents that autonomously handle customer service. If the pipelines that feed those agents don't deliver the specific data relevant to an individual customer and their problem, the agent won't be effective.
Semantic modeling -- ensuring that metadata is consistently and clearly classified whenever it is ingested or transformed -- is one means of improving search relevance and is gaining popularity as enterprises invest more heavily in AI initiatives. Graphwise's knowledge graph serves as a semantic layer, adding context to data and finding relationships between data points to make them easily discoverable.
GraphRAG unites large language models, an enterprise's data, a structured knowledge graph and multiple search methods such as similarity search and keyword search to deliver appropriate data to agents and other applications to provide accurate outputs at a much higher rate than standard RAG.
"The development of GraphRAG was driven by a specific structural failure in the market we call the 'Prototype Plateau,'" said Andreas Blumauer, founder of Semantic Web Company and Graphwise's senior vice president of growth. "While customers were indeed requesting better accuracy, the primary motivation came from observing enterprises stuck in a cycle of failed pilots."
In particular, RAG systems didn't provide enough context to retrieving data, he continued.
"Our motivation was to transform RAG from a simple associative engine into a reasoning engine," Blumauer said. "By injecting a 'Semantic Backbone', we moved beyond probability-based guesses to explicit, logic-based relationships."
Specific GraphRAG features include the following:
Semantic Metadata Control Plane, a semantic model designed to substantially improve the accuracy of AI outputs, including reducing the likelihood of AI hallucinations, by grounding responses in an enterprise's consistent metadata.
Explainability and Provenance Panels that display how AI responses are generated, enabling users to check for accuracy and supporting regulatory compliance by providing transparency.
Visual debugging and monitoring capabilities that allow developers and engineers to trace an error path and drastically reduce the amount of time previously needed to troubleshoot.
A low-code interface that enables business users to adjust AI logic without involving Python code experts.
Built-in templates that provide governance and enable query expansion that would otherwise require extensive research and development and technical support.
Simple Knowledge Organization System (SKOS)-like enrichment to capture domain-specific intelligence so that AI tools can understand an enterprise's unique terminology and ensure that users get accurate responses regardless of how they phrase a query.
Most valuable are the Semantic Metadata Control Plane and SKOS-style enrichment, according to Morrison, who noted that the control plane is where enterprises can make data accessible and discoverable across their entire data estate while SKOS-style enrichment allows non-technical users to work with data.
Catanzano likewise highlighted the Semantic Metadata Control Plane. In addition, he noted the value of the low-code interface. Meanwhile, capabilities such as GraphRAG help Graphwise differentiate from competing graph technology vendors by integrating knowledge graphs with AI, Catanzano continued.
"Its capabilities, such as explainability, provenance, and domain-specific intelligence, position it as a leader in making generative AI reliable and scalable, surpassing the limitations of traditional graph database vendors," he said.
Looking ahead
With GraphRAG now available, Graphwise's product development plans include adding AI-assisted automation capabilities and improving the memory of its platform, according to Blumauer.
Memory initiatives include moving beyond session-based interactions to retaining user preferences and context to provide more personalization. AI-assisted automation plans include tools that augment text with metadata to make it discoverable and generating schemas to aid data modeling.
"A major theme is reducing manual effort through AI-assisted automation," Blumauer said.
Morrison advised that Graphwise to develop a multi-layer context graph to expand on the context GraphRAG currently provides.
Catanzano, meanwhile, suggested that Graphwise develop new integrations with data and AI providers to expand its ecosystem and create prebuilt templates that simplify its platform for enterprises in certain industries.
"Industry-specific templates … would not only deepen its value for current users but also attract new customers seeking tailored, ready-to-deploy solutions," he said.
Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than three decades of experience. He covers analytics and data management.
