Google

When a Large Language Model generates a specific string formatted as a tool call, it triggers an execution process. The model runs the designated tool, such as a web search, with the provided query. The output from this tool then replaces the original tool call string in the model's context. This newly inserted information is used by the LLM in subsequent prediction steps to formulate a correct and contextually aware answer.

Executing Tool Calls and Integrating Results in LLMs

After an external tool, such as a web search, returns the required information, the Large Language Model processes this data to formulate a final answer. This output can range from a direct statement, like 'So the answer is: Los Angeles', to a more complete, user-friendly sentence, such as 'The 2028 Olympics will be held in Los Angeles.'

Example of a Final LLM-Generated Answer Using a Tool

A large language model is asked: 'What is the capital of Australia and what is its population?' The model begins generating its response and produces the following internal state:

`The capital of Australia is Canberra. To find the population, I will search online. {tool: web-search, query: "population of Canberra"}`

The web search tool then executes and returns the text: `The population of Canberra in 2023 was 467,194.`

What will the model's internal context be immediately after this result i

A user asks a large language model, 'What is the current weather in Paris?' The model's internal process generates a command to call a weather tool, but the tool returns the following text: `API_LIMIT_EXCEEDED`. The model must now generate a final response to the user. Based on your understanding of how tool results are integrated, what kind of response should the model logically produce, and why?

Analyzing a Tool Execution Failure

A language model is tasked with answering a question that requires external information. Below are four distinct internal actions that occur. Arrange these actions in the correct chronological sequence from first to last.

Learn Before

Related