Skip to main content

How to add ad-hoc tool calling capability to LLMs and Chat Models

Prerequisites

This guide assumes familiarity with the following concepts:

In this guide we’ll build a Chain that does not rely on any special model APIs (like tool calling, which we showed in the Quickstart) and instead just prompts the model directly to invoke tools.

Setup

We’ll need to install the following packages:

yarn add @langchain/core zod

Set environment variables

# Optional, use LangSmith for best-in-class observability
LANGSMITH_API_KEY=your-api-key
LANGCHAIN_TRACING_V2=true

Create a tool

First, we need to create a tool to call. For this example, we will create a custom tool from a function. For more information on all details related to creating custom tools, please see this guide.

import { StructuredTool } from "@langchain/core/tools";
import { z } from "zod";

class Multiply extends StructuredTool {
schema = z.object({
first_int: z.number(),
second_int: z.number(),
});

name = "multiply";

description = "Multiply two integers together.";

async _call(input: z.infer<typeof this.schema>) {
return (input.first_int * input.second_int).toString();
}
}

const multiply = new Multiply();
console.log(multiply.name);
console.log(multiply.description);
multiply
Multiply two integers together.
await multiply.invoke({ first_int: 4, second_int: 5 });
20

Creating our prompt

We’ll want to write a prompt that specifies the tools the model has access to, the arguments to those tools, and the desired output format of the model. In this case we’ll instruct it to output a JSON blob of the form {"name": "...", "arguments": {...}}.

tip

As of langchain version 0.2.8, the renderTextDescription function now supports OpenAI-formatted tools.

import { renderTextDescription } from "langchain/tools/render";

const renderedTools = renderTextDescription([multiply]);
import { ChatPromptTemplate } from "@langchain/core/prompts";

const systemPrompt = `You are an assistant that has access to the following set of tools. Here are the names and descriptions for each tool:

{rendered_tools}

Given the user input, return the name and input of the tool to use. Return your response as a JSON blob with 'name' and 'arguments' keys.`;

const prompt = ChatPromptTemplate.fromMessages([
["system", systemPrompt],
["user", "{input}"],
]);

Adding an output parser

We’ll use the JsonOutputParser for parsing our models output to JSON.

Pick your chat model:

Install dependencies

yarn add @langchain/openai 

Add environment variables

OPENAI_API_KEY=your-api-key

Instantiate the model

import { ChatOpenAI } from "@langchain/openai";

const model = new ChatOpenAI({
model: "gpt-3.5-turbo",
temperature: 0
});
import { JsonOutputParser } from "@langchain/core/output_parsers";
const chain = prompt.pipe(model).pipe(new JsonOutputParser());
await chain.invoke({
input: "what's thirteen times 4",
rendered_tools: renderedTools,
});
{ name: 'multiply', arguments: [ 13, 4 ] }

Invoking the tool

We can invoke the tool as part of the chain by passing along the model-generated “arguments” to it:

import { RunnableLambda, RunnablePick } from "@langchain/core/runnables";

const chain = prompt
.pipe(model)
.pipe(new JsonOutputParser())
.pipe(new RunnablePick("arguments"))
.pipe(
new RunnableLambda({
func: (input) =>
multiply.invoke({
first_int: input[0],
second_int: input[1],
}),
})
);
await chain.invoke({
input: "what's thirteen times 4",
rendered_tools: renderedTools,
});
52

Choosing from multiple tools

Suppose we have multiple tools we want the chain to be able to choose from:

class Add extends StructuredTool {
schema = z.object({
first_int: z.number(),
second_int: z.number(),
});

name = "add";

description = "Add two integers together.";

async _call(input: z.infer<typeof this.schema>) {
return (input.first_int + input.second_int).toString();
}
}
const add = new Add();

class Exponentiate extends StructuredTool {
schema = z.object({
first_int: z.number(),
second_int: z.number(),
});

name = "exponentiate";

description = "Exponentiate the base to the exponent power.";

async _call(input: z.infer<typeof this.schema>) {
return Math.pow(input.first_int, input.second_int).toString();
}
}
const exponentiate = new Exponentiate();

With function calling, we can do this like so:

If we want to run the model selected tool, we can do so using a function that returns the tool based on the model output. Specifically, our function will action return it’s own subchain that gets the “arguments” part of the model output and passes it to the chosen tool:

import { StructuredToolInterface } from "@langchain/core/tools";

const tools = [add, exponentiate, multiply];

const toolChain = (modelOutput) => {
const toolMap: Record<string, StructuredToolInterface> = Object.fromEntries(
tools.map((tool) => [tool.name, tool])
);
const chosenTool = toolMap[modelOutput.name];
return new RunnablePick("arguments").pipe(
new RunnableLambda({
func: (input) =>
chosenTool.invoke({
first_int: input[0],
second_int: input[1],
}),
})
);
};
const toolChainRunnable = new RunnableLambda({
func: toolChain,
});

const renderedTools = renderTextDescription(tools);
const systemPrompt = `You are an assistant that has access to the following set of tools. Here are the names and descriptions for each tool:

{rendered_tools}

Given the user input, return the name and input of the tool to use. Return your response as a JSON blob with 'name' and 'arguments' keys.`;

const prompt = ChatPromptTemplate.fromMessages([
["system", systemPrompt],
["user", "{input}"],
]);
const chain = prompt
.pipe(model)
.pipe(new JsonOutputParser())
.pipe(toolChainRunnable);
await chain.invoke({
input: "what's 3 plus 1132",
rendered_tools: renderedTools,
});
1135

Returning tool inputs

It can be helpful to return not only tool outputs but also tool inputs. We can easily do this with LCEL by RunnablePassthrough.assign-ing the tool output. This will take whatever the input is to the RunnablePassrthrough components (assumed to be a dictionary) and add a key to it while still passing through everything that’s currently in the input:

import { RunnablePassthrough } from "@langchain/core/runnables";

const chain = prompt
.pipe(model)
.pipe(new JsonOutputParser())
.pipe(RunnablePassthrough.assign({ output: toolChainRunnable }));
await chain.invoke({
input: "what's 3 plus 1132",
rendered_tools: renderedTools,
});
{ name: 'add', arguments: [ 3, 1132 ], output: '1135' }

What’s next?

This how-to guide shows the “happy path” when the model correctly outputs all the required tool information.

In reality, if you’re using more complex tools, you will start encountering errors from the model, especially for models that have not been fine tuned for tool calling and for less capable models.

You will need to be prepared to add strategies to improve the output from the model; e.g.,

  • Provide few shot examples.
  • Add error handling (e.g., catch the exception and feed it back to the LLM to ask it to correct its previous output).

Was this page helpful?


You can also leave detailed feedback on GitHub.