Incorporating tools into workflows

Now that we know how to create and use tools, let’s discuss how we can incorporate the tool-calling paradigm deeper into the workflows we’re developing.

Controlled generation

In Chapter 3, we started to discuss a controlled generation, when you want an LLM to follow a specific schema. We can improve our parsing workflows not only by creating more sophisticated and reliable parsers but also by being more strict in forcing an LLM to adhere to a certain schema. Calling a tool requires controlled generation since the generated payload should follow a specific schema, but we can take a step back and substitute our expected schema with a forced tool calling that follows the expected schema. LangChain has a built-in mechanism to help with that—an LLM has the with_structured_output method that takes a schema as a Pydantic model, converts it to a tool, invokes the LLM with a given prompt by forcing it to call this tool, and parses...