Tool Calling in LangChain | Generative AI using LangChain | Video 17 | CampusX

Summary

The video explains how LangChain connects language models to tools and how a program can use tool results to answer questions. The instructor demonstrates the process with a simple multiplication example, then builds a real-time currency-conversion application.

Key concepts

• Tools are Python functions made available to an LLM so it can request actions beyond generating text, such as searching the web, running commands, or calling an API.
• Tool binding registers tools with an LLM. A tool’s name, description, and input schema help the model determine what it can use and what arguments each tool expects. In the example, the instructor creates a multiplication tool and binds it using bind_tools.
• Tool calling is the model’s structured request to use a tool. The response may contain a tool name and arguments rather than ordinary text. A greeting does not require a tool, while a multiplication question can prompt the model to request the multiplication tool.
• Tool calling is not tool execution. The LLM recommends a tool and supplies arguments; the application code or programmer executes the function. This keeps execution under program control.
• Tool execution runs the requested function with the supplied arguments. LangChain returns the result as a tool message, which can be added to the conversation history and sent back to the LLM for a final natural-language response.

Practical walkthroughs

1. Multiplication tool

• The instructor defines a tool that multiplies two integers.
• After the tool is bound to the LLM, a multiplication request produces a tool call.
• The code executes the tool, appends the human, AI, and tool messages to the conversation history, and invokes the LLM again to produce the final answer.

2. Real-time currency conversion

The application uses two tools:

• One calls an exchange-rate API to retrieve a rate for a base and target currency.
• The other multiplies an amount by the retrieved rate.

The instructor demonstrates why asking the LLM to call both tools at once can fail: it may guess a conversion rate from its training data before the API tool has returned the current rate.

To prevent this, the conversion-rate input is marked as an injected tool argument. The LLM supplies the amount, while the program inserts the rate after executing the API tool. The code then runs the tools in sequence, adds both tool results to the message history, and asks the LLM to formulate the final response. The API requires an API key.

Is the example an AI agent?

The instructor says the currency-conversion application, as implemented, is not yet an AI agent. The programmer controls the execution flow and decides how the tools are run. An agent would autonomously determine the steps, call the rate tool, use its result to call the conversion tool, and return the answer. The video presents this distinction as preparation for a later lesson on building agents.

Guides and tutorials covered

• A step-by-step tutorial on creating, binding, calling, and executing LangChain tools.
• A practical walkthrough of a currency-conversion application using an exchange-rate API.
• An explanation of injected tool arguments and why they help enforce sequential tool use.
• A comparison between a manually orchestrated tool workflow and an autonomous AI agent.

Main speaker/source: Nitesh, the video’s instructor from CampusX.