ReACT pattern

What we did before is a sequence of steps:

1. We call a model
2. It decides to call one of the available tools (a calculator) and generates a payload (input) for this tool
3. We call the tool and give results back to the model (please note that we send the whole history of our interactions with an LLM since it doesn’t cache it server-side, although sometimes it’s done client-side on the SDK level)
4. These steps continue until the model decides to generate a final output that can be passed to a user

In other words, we continuously asked an LLM to reason and then we acted on this reasoning. It is a so-called ReACT pattern. ReACT stands for Reason+Act, and this pattern was introduced in the famous paper published by Google Search and Princeton University (the project resulted from an internship at Google) [7].

Figure 9.1 – ReACT pattern

Let’s look at a very simple example:

react_prompt = ChatPromptTemplate...