Language Model Middleware

Learn how to use middleware to enhance the behavior of language models

Language Model Middleware

Language model middleware is a way to enhance the behavior of language models by intercepting and modifying the calls to the language model.

It can be used to add features like guardrails, RAG, caching, and logging in a language model agnostic way. Such middleware can be developed and distributed independently from the language models that they are applied to.

Using Language Model Middleware

You can use language model middleware with the wrapLanguageModel function. It takes a language model and a language model middleware and returns a new language model that incorporates the middleware.

import { wrapLanguageModel, streamText } from 'ai';

const wrappedLanguageModel = wrapLanguageModel({
  model: yourModel,
  middleware: yourLanguageModelMiddleware,
});

The wrapped language model can be used just like any other language model, e.g. in streamText:

const result = streamText({
  model: wrappedLanguageModel,
  prompt: 'What cities are in the United States?',
});

Multiple Middlewares

You can provide multiple middlewares to the wrapLanguageModel function. The middlewares will be applied in the order they are provided.

const wrappedLanguageModel = wrapLanguageModel({
  model: yourModel,
  middleware: [firstMiddleware, secondMiddleware],
});

// applied as: firstMiddleware(secondMiddleware(yourModel))

Built-in Middleware

The AI SDK comes with several built-in middlewares that you can use to configure language models:

• extractReasoningMiddleware: Extracts reasoning information from the generated text and exposes it as a reasoning property on the result.
• extractJsonMiddleware: Extracts JSON from text content by stripping markdown code fences. Useful when using Output.object() with models that wrap JSON responses in code blocks.
• simulateStreamingMiddleware: Simulates streaming behavior with responses from non-streaming language models.
• defaultSettingsMiddleware: Applies default settings to a language model.
• addToolInputExamplesMiddleware: Adds tool input examples to tool descriptions for providers that don’t natively support the inputExamples property.

Extract Reasoning

Some providers and models expose reasoning information in the generated text using special tags, e.g. <think> and </think>.

The extractReasoningMiddleware function can be used to extract this reasoning information and expose it as a reasoning property on the result.

import { wrapLanguageModel, extractReasoningMiddleware } from 'ai';

const model = wrapLanguageModel({
  model: yourModel,
  middleware: extractReasoningMiddleware({ tagName: 'think' }),
});

You can then use that enhanced model in functions like generateText and streamText.

The extractReasoningMiddleware function also includes a startWithReasoning option. When set to true, the reasoning tag will be prepended to the generated text. This is useful for models that do not include the reasoning tag at the beginning of the response.

Extract JSON

Some models wrap JSON responses in markdown code fences (e.g., ```json ... ```) even when you request structured output.

The extractJsonMiddleware function strips these code fences from the response, making it compatible with Output.object().

import {
  wrapLanguageModel,
  extractJsonMiddleware,
  Output,
  generateText,
} from 'ai';
import { z } from 'zod';

const model = wrapLanguageModel({
  model: yourModel,
  middleware: extractJsonMiddleware(),
});

const result = await generateText({
  model,
  output: Output.object({
    schema: z.object({
      name: z.string(),
      ingredients: z.array(z.string()),
    }),
  }),
  prompt: 'Generate a recipe.',
});

You can also provide a custom transform function for models that use different formatting:

const model = wrapLanguageModel({
  model: yourModel,
  middleware: extractJsonMiddleware({
    transform: text => text.replace(/^PREFIX/, '').replace(/SUFFIX$/, ''),
  }),
});

Simulate Streaming

The simulateStreamingMiddleware function can be used to simulate streaming behavior with responses from non-streaming language models. This is useful when you want to maintain a consistent streaming interface even when using models that only provide complete responses.

import { wrapLanguageModel, simulateStreamingMiddleware } from 'ai';

const model = wrapLanguageModel({
  model: yourModel,
  middleware: simulateStreamingMiddleware(),
});

Default Settings

The defaultSettingsMiddleware function can be used to apply default settings to a language model.

import { wrapLanguageModel, defaultSettingsMiddleware } from 'ai';

const model = wrapLanguageModel({
  model: yourModel,
  middleware: defaultSettingsMiddleware({
    settings: {
      temperature: 0.5,
      maxOutputTokens: 800,
      providerOptions: { openai: { store: false } },
    },
  }),
});

Add Tool Input Examples

The addToolInputExamplesMiddleware function adds tool input examples to tool descriptions. This is useful for providers that don’t natively support the inputExamples property on tools. The middleware serializes the examples into the tool’s description text so models can still benefit from seeing example inputs.

import { wrapLanguageModel, addToolInputExamplesMiddleware } from 'ai';

const model = wrapLanguageModel({
  model: yourModel,
  middleware: addToolInputExamplesMiddleware({
    prefix: 'Input Examples:',
  }),
});

When you define a tool with inputExamples, the middleware will append them to the tool’s description:

import { generateText, tool } from 'ai';
import { z } from 'zod';

const result = await generateText({
  model, // wrapped model from above
  tools: {
    weather: tool({
      description: 'Get the weather in a location',
      inputSchema: z.object({
        location: z.string(),
      }),
      inputExamples: [
        { input: { location: 'San Francisco' } },
        { input: { location: 'London' } },
      ],
    }),
  },
  prompt: 'What is the weather in Tokyo?',
});

The tool description will be transformed to:

Get the weather in a location

Input Examples:
{"location":"San Francisco"}
{"location":"London"}

Options

• prefix (optional): A prefix text to prepend before the examples. Default: 'Input Examples:'.
• format (optional): A custom formatter function for each example. Receives the example object and its index. Default: JSON.stringify(example.input).
• remove (optional): Whether to remove the inputExamples property from the tool after adding them to the description. Default: true.

Implementing Language Model Middleware

Implementing language model middleware is advanced functionality and requires a solid understanding of the language model specification.

You can implement any of the following three functions to modify the behavior of the language model:

1. transformParams: Transforms the parameters before they are passed to the language model, for both doGenerate and doStream.
2. wrapGenerate: Wraps the doGenerate method of the language model. You can modify the parameters, call the language model, and modify the result.
3. wrapStream: Wraps the doStream method of the language model. You can modify the parameters, call the language model, and modify the result.

Here are some examples of how to implement language model middleware:

Examples

These examples are not meant to be used in production. They are just to show how you can use middleware to enhance the behavior of language models.

Logging

This example shows how to log the parameters and generated text of a language model call.

import type {
  LanguageModelV3Middleware,
  LanguageModelV3StreamPart,
} from '@ai-sdk/provider';

export const yourLogMiddleware: LanguageModelV3Middleware = {
  wrapGenerate: async ({ doGenerate, params }) => {
    console.log('doGenerate called');
    console.log(`params: ${JSON.stringify(params, null, 2)}`);

    const result = await doGenerate();

    console.log('doGenerate finished');
    console.log(`generated text: ${result.text}`);

    return result;
  },

  wrapStream: async ({ doStream, params }) => {
    console.log('doStream called');
    console.log(`params: ${JSON.stringify(params, null, 2)}`);

    const { stream, ...rest } = await doStream();

    let generatedText = '';
    const textBlocks = new Map<string, string>();

    const transformStream = new TransformStream<
      LanguageModelV3StreamPart,
      LanguageModelV3StreamPart
    >({
      transform(chunk, controller) {
        switch (chunk.type) {
          case 'text-start': {
            textBlocks.set(chunk.id, '');
            break;
          }
          case 'text-delta': {
            const existing = textBlocks.get(chunk.id) || '';
            textBlocks.set(chunk.id, existing + chunk.delta);
            generatedText += chunk.delta;
            break;
          }
          case 'text-end': {
            console.log(
              `Text block ${chunk.id} completed:`,
              textBlocks.get(chunk.id),
            );
            break;
          }
        }

        controller.enqueue(chunk);
      },

      flush() {
        console.log('doStream finished');
        console.log(`generated text: ${generatedText}`);
      },
    });

    return {
      stream: stream.pipeThrough(transformStream),
      ...rest,
    };
  },
};

Caching

This example shows how to build a simple cache for the generated text of a language model call.

import type { LanguageModelV3Middleware } from '@ai-sdk/provider';

const cache = new Map<string, any>();

export const yourCacheMiddleware: LanguageModelV3Middleware = {
  wrapGenerate: async ({ doGenerate, params }) => {
    const cacheKey = JSON.stringify(params);

    if (cache.has(cacheKey)) {
      return cache.get(cacheKey);
    }

    const result = await doGenerate();

    cache.set(cacheKey, result);

    return result;
  },

  // here you would implement the caching logic for streaming
};

Retrieval Augmented Generation (RAG)

This example shows how to use RAG as middleware.

Helper functions like getLastUserMessageText and findSources are not part of the AI SDK. They are just used in this example to illustrate the concept of RAG.

import type { LanguageModelV3Middleware } from '@ai-sdk/provider';

export const yourRagMiddleware: LanguageModelV3Middleware = {
  transformParams: async ({ params }) => {
    const lastUserMessageText = getLastUserMessageText({
      prompt: params.prompt,
    });

    if (lastUserMessageText == null) {
      return params; // do not use RAG (send unmodified parameters)
    }

    const instruction =
      'Use the following information to answer the question:\n' +
      findSources({ text: lastUserMessageText })
        .map(chunk => JSON.stringify(chunk))
        .join('\n');

    return addToLastUserMessage({ params, text: instruction });
  },
};

Guardrails

Guard rails are a way to ensure that the generated text of a language model call is safe and appropriate. This example shows how to use guardrails as middleware.

import type { LanguageModelV3Middleware } from '@ai-sdk/provider';

export const yourGuardrailMiddleware: LanguageModelV3Middleware = {
  wrapGenerate: async ({ doGenerate }) => {
    const { text, ...rest } = await doGenerate();

    // filtering approach, e.g. for PII or other sensitive information:
    const cleanedText = text?.replace(/badword/g, '<REDACTED>');

    return { text: cleanedText, ...rest };
  },

  // here you would implement the guardrail logic for streaming
  // Note: streaming guardrails are difficult to implement, because
  // you do not know the full content of the stream until it's finished.
};

Configuring Per Request Custom Metadata

To send and access custom metadata in Middleware, you can use providerOptions. This is useful when building logging middleware where you want to pass additional context like user IDs, timestamps, or other contextual data that can help with tracking and debugging.

import { generateText, wrapLanguageModel } from 'ai';
import { openai } from '@ai-sdk/openai';
import type { LanguageModelV3Middleware } from '@ai-sdk/provider';

export const yourLogMiddleware: LanguageModelV3Middleware = {
  wrapGenerate: async ({ doGenerate, params }) => {
    console.log('METADATA', params?.providerMetadata?.yourLogMiddleware);
    const result = await doGenerate();
    return result;
  },
};

const { text } = await generateText({
  model: wrapLanguageModel({
    model: openai('gpt-4'),
    middleware: yourLogMiddleware,
  }),
  prompt: 'Invent a new holiday and describe its traditions.',
  providerOptions: {
    yourLogMiddleware: {
      hello: 'world',
    },
  },
});

console.log(text);