Gemini Deep Research

An agentic system

To build Deep Research, we developed a new planning system that enables the Gemini app to work through complex problems. For Deep Research, we trained Gemini models to be capable of:

• Breaking down the problem: When presented with a complex user query, the system first formulates a detailed research plan, breaking the problem into a series of smaller, manageable sub-tasks. You're in control of the plan: Gemini presents it to you, and you can refine it to make sure that it's focused on the right areas.

• Research: The model oversees the execution of this plan, and intelligently determines which sub-tasks can be tackled simultaneously and which need to be done sequentially. The model can use tools like search and web browsing to fetch information and reason over it. At each step, the model reasons over information available to decide its next move. We introduced a thinking panel for users to follow what the model has learned so far and what it intends to do next.

• Synthesis: Once the model determines that enough information has been gathered, it synthesises the findings into a comprehensive report. In building the report, Gemini critically evaluates the information, identifies key themes and inconsistencies, and structures the report in a logical and informative way, even performing multiple passes of self-critique to enhance clarity and detail.

New category, new problems, new solutions

In building Deep Research, we had to work through three significant technical challenges:

Multi-step planning

Research tasks require multiple steps of iterative planning. At each step, the model has to ground itself on all information gathered so far, then identify missing information and discrepancies that it wants to explore – all while trading off comprehensiveness with compute time and user-waiting time. Training the model to be effective at long multi-step planning in a data-efficient manner enabled us to make Deep Research function in an open domain setting across all topics.

Long-running inference

A typical Deep Research task involves many model calls over several minutes. This creates a challenge for building agents: It has to be built so that a single failure doesn't mean having to restart the task from the beginning.

To address this, we developed a novel asynchronous task manager that maintains a shared state between the planner and task models, allowing for graceful error recovery without restarting the entire task. This system is truly asynchronous: you can hop to a different app or quite literally turn off your computer after starting a Deep Research project and the next time that you visit Gemini, you'll get notified when your research is done.

Context management

Over the course of a research session, Gemini can process hundreds of pages of content. To maintain continuity and enable follow-up questions, we use Gemini's industry-leading 1 million token context window complemented with a RAG setup. This effectively allows the system to 'remember' everything that it has learned during that chat session, making it smarter the longer that you interact with it.

Now powered by 2.0 Flash Thinking (experimental)

When Deep Research launched in December, it was powered by Gemini 1.5 Pro. With the introduction of Gemini 2.0 Flash Thinking (experimental), we were able to dramatically improve both the quality and serving efficiency of this product. With thinking models, Gemini takes more time to plan out its approach before it makes its next steps. This innate characteristic of self-reflection and planning makes it a great fit for these kinds of long-running agentic tasks. What we see is that Gemini is now even better at all stages of research and delivers more detailed reports. At the same time, the compute efficiency of the Flash model allows us to expand access to Deep Research to far more users. We're really excited about developing Flash and Thinking models in general, and expect Deep Research to keep getting better and better.

What's next

We built the system to be versatile, so that over time, we can expand its capabilities by giving you more control over what it can browse and giving it sources beyond the open web.

We are excited to see how people use Deep Research, and these real-world experiences will inform how we continue to build and improve Deep Research. Ultimately, our goal is a truly agentic and universally helpful AI assistant.