Retrieval augmented generation (RAG) is not enough for chatbot accuracy – it needs knowledge graphs, according to Illumex, which has revealed more about its knowledge graphs used for context-blind AI large language models.
Israeli startup Illumex wants to bring database item relationships and data in structured, block-based storage to GenAI large language model (LLM) chatbots. Up until now they have been trained using unstructured (file and object) text, audio, image and video data. However much mission-critical business data is stored in relational databases with metadata – data contents are connected and interdependent. Knowledge graphs model how pairs of items (entities) are related, and such relationships can be used to improve chatbot accuracy and relevance. Illumex’s General Semantic Fabric (GSF) aims to be a mechanism for achieving this.
Founder and CEO Inna Tokarev Sela explained in a canned quote: “Our platform is designed to bridge the gap between enterprise data complexity and the transformative power of generative AI. By providing a turnkey solution that automates the complex process of mapping data semantics and resolving terminological inconsistencies across business silos, we’re enabling organizations to achieve true AI and GenAI readiness.”
A Gartner report, “Cool Vendors in Data Management: GenAI Disrupts Traditional Technologies,” suggests data management leaders should: “Deliver customized, context-aware and more accurate GenAI results by deploying RAG, knowledge graphs and other semantic technologies to leverage your organization’s existing data. Context-aware results are especially helpful for providing concrete guidance to users who engage with data via natural language queries.”
Knowledge graphs store and model relationships between entities (events, objects, concepts, or situations), between so-called head and tail entities, with a “triple” referring to a head entity + relationship + tail entity. Such triples can be linked and the relationships are the semantics.
The Gartner report explains: “In our opinion, illumex’s platform directly addresses this need by:
- Automating the creation of semantically enriched knowledge graphs
- Mapping enterprise data to industry-specific ontologies
- Enabling more precise and contextual natural language queries
- Providing an alternative to RAG that enables transparent and governed data interactions with LLMs”
LLMs use semantic search based on finding vectors (encoded tokens describing abstracted as aspects of text, audio, image or video data) similar to a chatbot user’s input request. The trick Illumex wants to pull off is getting its separate knowledge graph semantics used as well. But you cannot simply vectorize a knowledge graph describing structured data as you can a piece of text. Instead you abstract its inherent relationships – Illumex’s GSF – and enable an LLM to process this by using a dedicated application.
An Illumex spokesperson told us: “Today’s LLMs do not yet use knowledge graphs directly, but unifying the two technologies is an active area of research (see here).
“Currently, LLMs are integrated with knowledge graphs in two ways:
- LLMs are used to automate the creation and maintenance of knowledge graphs.
- Agent-like applications enable LLMs to retrieve information from knowledge graphs and use it to augment answers (the integration is in the application layer and not directly within the LLMs).
“Illumex uses both of these approaches, integrating custom knowledge graphs with any of the commercially available LLMs.”
We understand that these agent-like applications are like connectors between Illumex-created knowledge graphs and LLMs. Without these connectors Illumex and its knowledge graphs are stranded. These knowledge graph agents will be specific, at least initially, to a particular structured dataset’s knowledge graphs and a particular LLM. They’ll need to be created by Illumex itself or an Illumex customer wanting to use the GSF to improve the output from an LLM it uses or wants to use.
On being queried about this, Illumex told us: “Illumex has actually already created the connector themselves, which they call illumex Omni. Users can access a chat interface within their organization’s Slack. When the user asks a question, Omni interprets the question and maps it to the correct data objects using the knowledge graph, generates SQL and retrieves the relevant data, and then passes the context and data to an LLM to get an answer for the user. Omni supports open source LLMs, as well as commercial LLMs like ChatGPT.”
Check out an Illumex Omni video here.
Our understanding is that Illumex’s progress depends upon demonstrating that its GSF significantly and verifiably improves LLM responses in a RAG and vector-based semantic search environment that produces inadequate and/or wrong responses. Response accuracy and relevance should hopefully shoot up when its GSF is added to the mix.
Bootnote
The “Unifying Large Language Models and Knowledge Graphs: A Roadmap” research paper discuses how knowledge graphs can be used to improve LLMs, and LLMs provide help for knowledge graph creation. It also suggests a three-stage roadmap leading to a unification between LLMs and knowledge graphs.
