Building Reliable GenAI Systems by Harnessing the Power of LLMs and Ontological Knowledge Graphs

Recent research has shed light on a fundamental aspect of Large Language Models (LLMs) - hallucination is an inherent part of their architecture [1]. Just as a car's engine relies on controlled combustion of highly inflammable gasoline, LLMs have intrinsic strengths and weaknesses that must be carefully managed.

‍

The Intrinsic Nature of LLM Hallucinations

The paper, "LLMs Will Always Hallucinate, and We Need to Live with This" [1], establishes that hallucinations in LLMs are not merely occasional errors, but stem from the very mathematical and logical structure of these models [1]. No amount of architectural improvements, dataset enhancements, or fact-checking can fully eliminate this tendency.

However, this does not mean that LLMs are inherently dangerous. Just as we trust car engines despite the volatile nature of gasoline, we can harness the power of LLMs by integrating them into well-designed systems. A car is a complex system with many components working together to safely control the combustion process. Similarly, LLMs can be part of a larger GenAI system with additional components to mitigate their weaknesses.

‍

GenAI: A System, Not Just an LLM

The key is to view GenAI not merely as an LLM, but as a system composed of multiple components working in harmony. By combining LLMs with other elements such as ontological knowledge graphs, embeddings, rules, and logic, we can create a robust infrastructure for building reliable GenAI applications.

‍

The Power of Ontological Knowledge Graphs

Ontological knowledge graphs play a crucial role in this system. They bring conceptual structures and symbolic knowledge to bear on the LLM's tasks, providing a framework for grounding the model's outputs in real-world concepts and relationships.

This semantic grounding helps to mitigate the risks of hallucination. While the LLM may still generate inconsistencies at a local level, the ontological knowledge graph provides a global structure that can help to identify and correct these errors.

As Jérémy Ravenel points out in his post, ontologies can be thought of as "configuration files" for LLMs. They provide the necessary semantic scaffolding to guide the model's outputs, ensuring that they align with the desired conceptual framework. Just as a car's onboard computer uses sensors and configuration settings to control the engine, an ontological knowledge graph can steer an LLM towards more reliable and consistent outputs.

‍

Crafting, Managing, and Steering the System

Of course, building such a system is not a trivial task. It requires careful crafting, ongoing management, and active steering to ensure that the LLM's outputs align with the desired outcomes.

This means being fully cognizant of the possibilities of errors and having robust mechanisms in place to detect and handle them. It also means continually refining and updating the ontological knowledge graph to keep pace with the evolving needs of the application.

Just as a car requires regular maintenance, tuning, and upgrades to operate safely and efficiently, a GenAI system needs ongoing care and adjustment. The ontological knowledge graph must be curated and expanded, the LLM fine-tuned on new data, and the overall system architecture adapted to changing requirements.

‍

Conclusion

In conclusion, while LLM hallucinations may be an intrinsic part of their architecture, this does not preclude their use in reliable GenAI systems. By viewing GenAI as a system composed of multiple components, including ontological knowledge graphs, we can harness the power of LLMs while mitigating their weaknesses.

This is the approach pioneered by CogniSwitch, and it represents an exciting frontier in the development of GenAI applications. By using ontologies and connected enterprise knowledge as configuration files for the LLM, together with rules and logic that helps inferencing, we can steer these powerful models towards reliable, well-defined outcomes. As we continue to refine these systems and techniques, we look forward to a future where the power of LLMs are safely and reliably harnessed for a wide range of transformative applications.

Just as the automobile revolutionized transportation despite the dangers of the internal combustion engine, GenAI has the potential to transform many aspects of our lives and work. The key is to approach it with a systems mindset, leveraging the right components in building a powerful system that enables robust and reliable applications. With careful engineering and ongoing maintenance, we can harness the power of LLMs while keeping their hallucinations firmly in check.

‍

Sources:

[1] LLMs Will Always Hallucinate, and We Need to Live With This (Research Paper)

[2] URL to Jérémy’s post