Skip to main content

Using OpenAI functions

This walkthrough demonstrates how to incorporate OpenAI function-calling API's in a chain. We'll go over:

  1. How to use functions to get structured outputs from ChatOpenAI
  2. How to create a generic chain that uses (multiple) functions
  3. How to create a chain that actually executes the chosen function

Getting structured outputs​

We can take advantage of OpenAI functions to try and force the model to return a particular kind of structured output. We'll use createStructuredOutputRunnable to create our chain, which takes the desired structured output as a valid JSONSchema.

import { ChatPromptTemplate } from "@langchain/core/prompts";
import { ChatOpenAI } from "@langchain/openai";
import { JsonOutputFunctionsParser } from "langchain/output_parsers";

const jsonSchema = {
  title: "Person",
  description: "Identifying information about a person.",
  type: "object",
  properties: {
    name: { title: "Name", description: "The person's name", type: "string" },
    age: { title: "Age", description: "The person's age", type: "integer" },
    fav_food: {
      title: "Fav Food",
      description: "The person's favorite food",
      type: "string",
    },
  },
  required: ["name", "age"],
};

const model = new ChatOpenAI();
const prompt = ChatPromptTemplate.fromMessages([
  ["human", "Human description: {description}"],
]);
const outputParser = new JsonOutputFunctionsParser();

const runnable = prompt
  .pipe(model.withStructuredOutput(jsonSchema))
  .pipe(outputParser);

const response = await runnable.invoke({
  description:
    "My name's John Doe and I'm 30 years old. My favorite kind of food are chocolate chip cookies.",
});
console.log(response);
/*
  { name: 'John Doe', age: 30, fav_food: 'chocolate chip cookies' }
*/

API Reference:

tip

You can see a LangSmith trace of this example here.

Creating a generic OpenAI functions chain​

To create a generic OpenAI functions chain, we can use the createOpenaiFnRunnable method. This is the same as createStructuredOutputRunnable except that instead of taking a single output schema, it takes a sequence of function definitions.

import { ChatPromptTemplate } from "@langchain/core/prompts";
import { ChatOpenAI } from "@langchain/openai";
import { createOpenAIFnRunnable } from "langchain/chains/openai_functions";
import { JsonOutputFunctionsParser } from "langchain/output_parsers";

const openAIFunction = {
  name: "get_person_details",
  description: "Get details about a person",
  parameters: {
    title: "Person",
    description: "Identifying information about a person.",
    type: "object",
    properties: {
      name: { title: "Name", description: "The person's name", type: "string" },
      age: { title: "Age", description: "The person's age", type: "integer" },
      fav_food: {
        title: "Fav Food",
        description: "The person's favorite food",
        type: "string",
      },
    },
    required: ["name", "age"],
  },
};

const model = new ChatOpenAI();
const prompt = ChatPromptTemplate.fromMessages([
  ["human", "Human description: {description}"],
]);
const outputParser = new JsonOutputFunctionsParser();

const runnable = createOpenAIFnRunnable({
  functions: [openAIFunction],
  llm: model,
  prompt,
  enforceSingleFunctionUsage: true, // Default is true
  outputParser,
});
const response = await runnable.invoke({
  description:
    "My name's John Doe and I'm 30 years old. My favorite kind of food are chocolate chip cookies.",
});
console.log(response);
/*
  { name: 'John Doe', age: 30, fav_food: 'chocolate chip cookies' }
*/

API Reference:

tip

You can see a LangSmith trace of this example here.

Multiple functions​

import { ChatPromptTemplate } from "@langchain/core/prompts";
import { ChatOpenAI } from "@langchain/openai";
import { createOpenAIFnRunnable } from "langchain/chains/openai_functions";
import { JsonOutputFunctionsParser } from "langchain/output_parsers";

const personDetailsFunction = {
  name: "get_person_details",
  description: "Get details about a person",
  parameters: {
    title: "Person",
    description: "Identifying information about a person.",
    type: "object",
    properties: {
      name: { title: "Name", description: "The person's name", type: "string" },
      age: { title: "Age", description: "The person's age", type: "integer" },
      fav_food: {
        title: "Fav Food",
        description: "The person's favorite food",
        type: "string",
      },
    },
    required: ["name", "age"],
  },
};

const weatherFunction = {
  name: "get_weather",
  description: "Get the weather for a location",
  parameters: {
    title: "Location",
    description: "The location to get the weather for.",
    type: "object",
    properties: {
      state: {
        title: "State",
        description: "The location's state",
        type: "string",
      },
      city: {
        title: "City",
        description: "The location's city",
        type: "string",
      },
      zip_code: {
        title: "Zip Code",
        description: "The locations's zip code",
        type: "number",
      },
    },
    required: ["state", "city"],
  },
};

const model = new ChatOpenAI();
const prompt = ChatPromptTemplate.fromMessages([
  ["human", "Question: {question}"],
]);
const outputParser = new JsonOutputFunctionsParser();

const runnable = createOpenAIFnRunnable({
  functions: [personDetailsFunction, weatherFunction],
  llm: model,
  prompt,
  enforceSingleFunctionUsage: false, // Default is true
  outputParser,
});
const response = await runnable.invoke({
  question: "What's the weather like in Berkeley CA?",
});
console.log(response);
/*
  { state: 'CA', city: 'Berkeley' }
*/

API Reference:

tip

You can see a LangSmith trace of this example here.

Help us out by providing feedback on this documentation page: