Contextual AI Inc., an artificial intelligence development startup founded earlier this year, exited stealth mode today with $20 million in seed funding.

Palo Alto, California-based Contextual AI raised the capital from a group of investors led by Bain Capital Ventures. The investment firm was joined by Lightspeed Venture Partners, Greycroft and SV Angel. More than a half-dozen angel investors participated as well.

Contextual AI is led by Chief Executive Officer Douwe Kiela and Chief Technology Officer Amanpreet Singh. Before founding the startup, Kiela and Singh held senior research positions at Hugging Face Inc., a company with a popular platform for hosting AI models. They earlier worked at Meta Platforms Inc.’s machine learning research division.

Contextual AI focuses on building large language models for the enterprise market. According to the startup, its models are less prone to AI hallucinations than current neural networks. That means they’re less likely to make up information when responding to user queries.

For added measure, the startup is equipping its neural networks with the ability to cite the sources from which they retrieve information. That will make it easier for organizations to verify the accuracy of AI responses.

The language models Contextual AI is developing are based on a machine learning technique called retrieval augmented generation, or RAG. Besides reducing the risk of AI hallucinations, the startup says, its technology enables models to answer more types of questions. It does so by increasing the amount of data that a neural network can use when responding to a query.

Historically, language models could only generate answers based on the dataset on which they were trained. Supplying a model with additional information often requires retraining it, which is a time-consuming and expensive process. As a result, it’s often impractical to update an AI system’s built-in knowledge base.

Not refreshing a model’s knowledge base as new information becomes available can potentially lead to out-of-date answers. According to Contextual AI, its RAG technology addresses that challenge. The technology equips neural networks with a component that can augment their built-in knowledge base with information from external sources.

If a user asks a RAG-powered model to provide an overview of the latest iPhone lineup, it could retrieve relevant details from Apple Inc.’s website. In the enterprise, the technology can also be used to fetch records from internal systems such as databases. The result, according to Context AI, is that its models can be more easily customized to companies’ requirements.

Many large language models are delivered on a managed basis. Contextual AI, in contrast, intends to give customers the option of deploying its models on their own cloud infrastructure. That means the data that a company sends to the startup’s models doesn’t have to leave the corporate network.

Contextual AI disclosed today that several Fortune 500 are in talks to pilot its software. The startup, which currently has eight employees, reportedly plans to hire 12 more workers by year’s end to support its growth efforts. Contextual AI will also use a portion of its $20 million seed round to build an infrastructure environment for training large language models. 

Image: Unsplash