Meet the researcher fighting AI hallucinations at Google Cloud

For Cyrus Rashtchian, making AI more reliable is a real responsibility.

Cyrus is a senior research scientist at Google. With a background in theoretical computer science and a postdoctoral fellowship focused on computer science and engineering, he has spent years examining the mathematical boundaries of how algorithms classify information. Lately, his focus has shifted to large language models, specifically a persistent issue known as hallucination, which is when an artificial intelligence model lacks the correct facts but confidently invents an answer anyway.

“I’ve started to realize that there’s a big disconnect between what people think AI can do and what it can actually do,” Cyrus said.

This disconnect made him appreciate the massive responsibility to ensure Google’s AI models work well. If a model is only 80% accurate, then that means those who use it to build for themselves, or for their customers, will only ever be 80% accurate. “I don't think we need to settle for noisy, hallucinated, chaotic or non-helpful responses,” he told me.

How agentic RAG reduces AI hallucinations in enterprise search

To work on these hallucinated responses directly, Cyrus and his colleagues across Google Research and Google Cloud thought enterprise workflows would be a good place to start.

Currently, many businesses use a framework called retrieval-augmented generation (RAG) to connect AI models to their private databases. In a standard setup, a user asks a question, the system searches the database once, pulls a few relevant documents, and generates an answer. If the answer requires synthesizing information from three different documents across two different databases, the system struggles. It finds one piece of the puzzle, gives up on the rest, and attempts to guess the missing context. This is where gaps form.

To solve this, the team developed an approach called Agentic RAG on Gemini Enterprise Agent Platform. Instead of treating the AI as a single search engine, they structured it like an organized research department. The system breaks complex requests down into smaller pieces, routing targeted search queries to specific databases to find sufficient context to answer a user’s query.

The root of its success? It asks the model to evaluate its own findings.

“A big thing our team does is multi-stage prompting, where you’ll prompt the model to do something, but then you prompt the model again to check its work,” Cyrus said. This process, often referred to as “critique and revise,” forces the system to pause and evaluate its output.

Imagine you’re managing your store inventory. If you have search agents, they might pull data from what’s on your shelves and stop there. But what if there’s a miscategorization, like lookalike products or mislabeled items? With an agentic RAG system, it compares its findings against the original request (e.g., pull the top 10 SKUs), and forces the agent to go back and continue searching and evaluating until the context is sufficient for a definitive answer.

Teaching models to read between the lines

It’s helpful to think of this work in phases, or evolutions. Phase one is making sure AI works responsibly, without hallucinations. The team’s work on RAG and Agent Platform is a step forward in that direction.

Phase two is intent, which helps AI read between the lines to understand what the human is actually asking for. Intent is grounded in a field called pragmatics, a subfield of linguistics and philosophy that studies the context of language. Humans rarely spell out every literal detail of their intent, meaning we rely heavily on the unspoken context shared between the speaker and the listener. This is notoriously hard for an AI to get right.

“We’ve been talking about this area called pragmatics, which is this fancy word for what people mean when they don't say everything they mean,” Cyrus told me. “If you ask me, 'Are you in New York?' I know that means, 'Do you work out of the New York office?' You weren't just curious what state I'm in. There’s all this shared context that makes that question make sense. When it comes to LLMs, we need a definition for what the shared context is between the user and the system.”

Importantly, intent is all about helping AI better understand humans, not the other way around. When I asked if phases one and two meant educating people on using AI differently or better, perhaps ways to make it easier for AI to understand our intent, he said users shouldn’t have to change their behaviors.

“I think it's something we can solve,” he said. “The user shouldn't take on this responsibility. The AI should work, the system should work, things should be predictable, consistent, and have the right answer as best as possible.”

The third phase of fighting hallucinations

Long-term, the third phase of this work is about adaptivity. It’s impressive when AI can help you write a personalized email, but will it understand your work culture’s specific vibe, how you interact with colleagues, or the right time to be more formal with a director versus a peer?

“We’re trying to build more adaptivity into the models, and I think that can make them better. We also hope that if we're using this at work, it somehow embodies the work culture in some way, using the right language or cadence. I spent a lot of time thinking about phase one and two, but we’re building phase three at the same time.”

To learn more about this research, you can read the full research deep dive here. This new RAG feature is now available as a public preview offering in Gemini Enterprise Agent Platform, and it will serve as the basis for future agentic search offerings later in 2026.