Ontologies for Enabling Automated Graph LLM Workflows

In this webinar, Prasad Yalamanchi, founder and CEO of Lead Semantics, discusses the integration of ontologies and Large Language Models (LLMs) to enhance automated workflows. The focus is on leveraging ontologies to improve the context and accuracy of Retrieval Augmented Generation (RAG) solutions, addressing the common issue of LLMs generating contextually inappropriate language.

Prasad explains that digital workflows consist of a series of independent tasks, each with specific inputs and outputs. Automating these workflows requires a master system to coordinate between tasks, ensuring that the outputs of one task align with the inputs of the next. This coordination is facilitated by agents, which are autonomous software entities capable of executing tasks based on internal logic and asynchronous messaging.

Ontologies play a crucial role in this process by providing a shared, machine-understandable view of the domain of interest. Unlike traditional relational or object-oriented modeling, ontologies represent entities, their properties, and relationships in a flexible and dynamic manner. This allows for better alignment and communication between agents, enhancing the overall efficiency and accuracy of the workflow.

Prasad introduces TextDistil, a platform developed over four years that uses NLP, deep learning, and semantic technology to extract knowledge from unstructured text and populate an ontology. He demonstrates how ontologies can power AI agents, making them more predictable and adaptable. By aligning agent ontologies with a central master ontology, the system can dynamically manage semantic, structural, and syntactic differences, ensuring smooth communication and collaboration.

The webinar includes a detailed example of an ESG (Environmental, Social, and Governance) compliance monitoring workflow. This workflow uses AI agents and graph RAG components to analyze news items, retrieve relevant data, and assess compliance with investment guidelines. The agents, equipped with domain-specific ontologies, communicate through a central ontology component to produce accurate and actionable insights.

Prasad concludes by highlighting the benefits of using ontologies with LLMs, such as minimizing hallucinations and increasing domain specificity. He emphasizes that ontologies can guide LLM interactions without requiring extensive retraining, making them a valuable tool for enhancing the performance of AI agents in real-world applications.

The webinar ends with a brief Q&A session, addressing questions about the differences between prompting LLMs with a knowledge graph versus text and the proven effectiveness of ontologies in minimizing hallucinations.