-
Book Overview & Buying
-
Table Of Contents
DeepSeek in Practice
By :
Tools are one of the most critical parts of building agents. Tool calling is what allows agents to interact and take actions in their environment. However, it’s important to note that we didn’t always have tool calling. In the early stages of LLMs (which is not that long ago), we used simple prompts to allow the LLMs to interact with the outside world.
Here’s an example:
Question: What is the capital of France?
Thought: I should look this up on Wikipedia.
Action: wikipedia: France
PAUSE
Observation: France is a country. The capital is Paris.
Answer: The capital of France is Paris.
In the preceding example, the LLM outputs an action (e.g., wikipedia: France). The system (us) sees that, runs the search, and feeds the results back to the LLM. It was crude, but it worked. This was the core mechanism behind popular papers such as the ReAct paper from Google (https://react-lm.github.io/).
As things evolved, this approach seemed brittle, and LLM...