Nadia Makarevich

Don't Fight Hallucinations. Make Them Impossible

The Heatseeker AI chat answers data questions for marketers who make decisions with million-dollar budgets. Wrong answers or hallucinated numbers are not an option here, as you can imagine ;)

The fight against them (hallucinations, not marketers) was long and painful. We started with a "naive" approach, which we all tried at some point, I imagine: "Hey, AI, here's all of our data, extract what we need, make no mistakes." It kinda worked but also kinda didn't - the output, especially on more complicated questions, could vary significantly.

Then we tried various techniques of validating the answer before sending the result to the user. This greatly improved the correctness of the data, but made the code fragile (this policing involved a lot of regular expressions), the chat very slow because of too much back-and-forth, and worst of all - it still couldn't guarantee that the answer was actually correct. Yes, all the numbers were citable, but were they answering the question the user actually asked? No post-validation can tell you that.

It was time to look at the problem more holistically. Can we beat AI hallucinations and make them structurally and architecturally impossible? In this talk, I’ll share our journey and how we flipped everything upside down to do that ;)