Patrick Lewis

AI chatbots sound eerily convincing—until you ask them about a topic you know a lot about.

The underlying large language models powering chatbots are generalists, limiting their usefulness in specialized domains that demand higher precision. But there’s a technique sweeping through Silicon Valley that helps empower these models with domain specific knowledge.

This method, called retrieval augmented generation, or RAG for short, was first introduced in a 2020 paper co-authored by Patrick Lewis, now director of Machine Learning at enterprise AI startup Cohere. RAG allows AI models to answer queries by drawing on external texts, be it company documents or a news website. It can also reduce hallucinations, and even give model’s access to up-to-the-minute information. Over the last year, the technique has been adopted by Microsoft, Google, Amazon, and Nvidia.

Lewis says he’s dreamt of ways computers can help us understand and leverage knowledge since toying with language models as a chemistry student. “I'm still kind of chasing that nearly 10 years later,” he says. Since joining Cohere in 2022, Lewis has continued to expand RAG by creating AI systems that cite any external sources used, making it easier for humans to verify their output. Lewis says the hope is that in future, “the entire world could be with footnotes that you can investigate if you want.”

*Disclosure: Investors in Cohere include Salesforce, where TIME co-chair and owner Marc Benioff is CEO.