AWS ReAct prompting

ReAct prompting

With ReAct prompting, an LLM can combine reasoning and action.

According to Shunyu Yao's article "ReAct: Synergizing Reasoning and Acting in Language Models," models are often used for reasoning or for acting, but they are not always effectively used for both at the same ti

CoT prompting shows promise for an LLM to reason and generate actions for straightforward tasks, like mathematics.

However, the inability to update information or access external contexts with CoT prompting can lead to errors in the output, such as fact hallucination.

With a ReAct framework, LLMs can generate both reasoning traces and task-specific actions that are based on external tools, such as Wikipedia pages or SQL databases.

This external context leads to more accurate and reliable output.

Consider the following prompts.

In the first example, a calculator tool is provided to the LLM.

In the second example, the prompt adds a SQL database tool in order to query stock prices.

Example prompt and output providing the LLM with a calculator tool

Example using a SQL database to query stock prices

Now you can extend this to use a SQL database tool for querying stock prices (with daily prices for the stock "ABC").