Douwe Kiela, one of the original pioneers of retrieval-augmented generation (RAG), decided to address this challenge by founding Contextual AI in June of 2023. The company’s platform uses specialized RAG agents to reduce hallucinations for specialized, highly accurate, and trustworthy AI applications.

Rather than building its models from scratch, Contextual AI chose to fine-tune open-source models for it’s platform. This allows developers and researchers to focus on grounding and eliminating hallucinations rather than spending resources to build a model from scratch.

After extensive trials and evaluations, the team determined that Llama models offered the best performance in terms of groundedness—or the lack of hallucinations—across various use cases. Llama 3.1 8B proved to be especially responsive to fine-tuning, delivering excellent accuracy, groundedness, and cost-performance.

Llama also serves as the foundation for other models in the Contextual AI platform, including reward models such as LMUnit, which evaluates how well AI responses satisfy specific criteria, and the Groundedness Reward Model, a real-time fact-checking model that reviews AI responses for potential hallucinations.

“We were able to take the experience from building our other models on Llama and use them as primordial for the reinforcement learning for our Grounded Language Model.” says William Berrios, Researcher at Contextual AI.

"A big reason we like Llama models is how effectively they can be fine-tuned for reward modeling, according to both the broader literature and our own experiments," says Rajan Vivek, Researcher at Contextual AI. "This enables us to optimize our language models for fine-grained customer preferences and use-cases."

When these models work together as a single unified system, they deliver high levels of accuracy through a modular approach. Customers can use the end-to-end platform as-is no coding required, or they can leverage the individual components through the interface and APIs to customize their own system to their needs.

Most importantly, Google Cloud is highly scalable with high-performance GPUs, providing the computational power needed for fine-tuning Llama models. High throughput and latency accelerate training and performance. With dynamic scaling, teams can spend less time worrying about infrastructure and more time refining models that help customers do more with their data.

For enterprises with massive volumes of highly technical and often multimodal documents,