Skip to main content

Classify Text into Labels

Tagging means labeling a document with classes such as:

  • sentiment
  • language
  • style (formal, informal etc.)
  • covered topics
  • political tendency

Image description

Overview​

Tagging has a few components:

  • function: Like extraction, tagging uses functions to specify how the model should tag a document
  • schema: defines how we want to tag the document

Quickstart​

Let’s see a very straightforward example of how we can use tool calling for tagging in LangChain. We’ll use the .withStructuredOutput(), supported on selected chat models.

Pick your chat model:

Install dependencies

tip

See this section for general instructions on installing integration packages.

yarn add @langchain/openai

Add environment variables

OPENAI_API_KEY=your-api-key

Instantiate the model

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

const llm = new ChatOpenAI({
  model: "gpt-4o-mini",
  temperature: 0
});

Let’s specify a Zod schema with a few properties and their expected type in our schema.

import { ChatPromptTemplate } from "@langchain/core/prompts";
import { z } from "zod";

const taggingPrompt = ChatPromptTemplate.fromTemplate(
  `Extract the desired information from the following passage.

Only extract the properties mentioned in the 'Classification' function.

Passage:
{input}
`
);

const classificationSchema = z.object({
  sentiment: z.string().describe("The sentiment of the text"),
  aggressiveness: z
    .number()
    .int()
    .min(1)
    .max(10)
    .describe("How aggressive the text is on a scale from 1 to 10"),
  language: z.string().describe("The language the text is written in"),
});

// Name is optional, but gives the models more clues as to what your schema represents
const llmWihStructuredOutput = llm.withStructuredOutput(classificationSchema, {
  name: "extractor",
});
const prompt1 = await taggingPrompt.invoke({
  input:
    "Estoy increiblemente contento de haberte conocido! Creo que seremos muy buenos amigos!",
});
await llmWihStructuredOutput.invoke(prompt1);
{ sentiment: 'contento', aggressiveness: 1, language: 'es' }

As we can see in the example, it correctly interprets what we want.

The results vary so that we may get, for example, sentiments in different languages (‘positive’, ‘enojado’ etc.).

We will see how to control these results in the next section.

Finer control​

Careful schema definition gives us more control over the model’s output.

Specifically, we can define:

  • possible values for each property
  • description to make sure that the model understands the property
  • required properties to be returned

Let’s redeclare our Zod schema to control for each of the previously mentioned aspects using enums:

import { z } from "zod";

const classificationSchema2 = z.object({
  sentiment: z
    .enum(["happy", "neutral", "sad"])
    .describe("The sentiment of the text"),
  aggressiveness: z
    .number()
    .int()
    .min(1)
    .max(5)
    .describe(
      "describes how aggressive the statement is, the higher the number the more aggressive"
    ),
  language: z
    .enum(["spanish", "english", "french", "german", "italian"])
    .describe("The language the text is written in"),
});
const taggingPrompt2 = ChatPromptTemplate.fromTemplate(
  `Extract the desired information from the following passage.

Only extract the properties mentioned in the 'Classification' function.

Passage:
{input}
`
);

const llmWihStructuredOutput2 = llm.withStructuredOutput(
  classificationSchema2,
  { name: "extractor" }
);

Now the answers will be restricted in a way we expect!

const prompt2 = await taggingPrompt2.invoke({
  input:
    "Estoy increiblemente contento de haberte conocido! Creo que seremos muy buenos amigos!",
});
await llmWihStructuredOutput2.invoke(prompt2);
{ sentiment: 'happy', aggressiveness: 1, language: 'spanish' }
const prompt3 = await taggingPrompt2.invoke({
  input: "Estoy muy enojado con vos! Te voy a dar tu merecido!",
});
await llmWihStructuredOutput2.invoke(prompt3);
{ sentiment: 'sad', aggressiveness: 5, language: 'spanish' }
const prompt4 = await taggingPrompt2.invoke({
  input: "Weather is ok here, I can go outside without much more than a coat",
});
await llmWihStructuredOutput2.invoke(prompt4);
{ sentiment: 'neutral', aggressiveness: 1, language: 'english' }

The LangSmith trace lets us peek under the hood:

Was this page helpful?


You can also leave detailed feedback on GitHub.