What is the ReAct (Reason + Act) pattern?

ReAct (Reason + Act) is a pattern of interaction with large language models (LLMs) that combines reasoning (Reason) and action (Act) into a single process. Instead of simply generating an answer, the model first formulates intermediate reasoning and then performs an action, such as querying an external data source or performing a calculation, before continuing to reason and generate the final answer.

This allows models to solve complex tasks more effectively, especially those requiring multiple steps of logic and interaction with external tools.

An example of a simplified dialogue with ReAct:

User: What is 15 multiplied by 3?
Model: Let's first calculate 15 * 3. [Performs calculation]
Model: 15 multiplied by 3 equals 45.

Thus, the model not only responds but also reasons and acts sequentially, which increases the accuracy and flexibility of responses.