Display reasoning separately
Separate a model's internal-style <think> section from its user-facing answer.
AI Tools · Function
extractReasoningMiddleware
Extracts delimited reasoning from generated model output and exposes it separately from the main text. It supports XML-style tags and custom literal delimiters.
How it works
The middleware examines generated output in both streaming and non-streaming modes. It recognizes configured delimiters, removes the delimited reasoning from the main text, and emits reasoning separately; separator and startWithReasoning control how sections are interpreted and joined.
Syntax
extractReasoningMiddleware({ tagName, separator?, startWithReasoning? })Parameters
| Name | Type | Description |
|---|---|---|
tagName | string | { opening: string; closing: string } | An XML tag name without angle brackets, or an object containing literal opening and closing delimiters. |
separator | string | Separator used between reasoning and text sections; defaults to "\n". |
startWithReasoning | boolean | Whether the response starts with reasoning tokens; defaults to false. |
Returns
A middleware object that processes streaming and non-streaming responses, extracts delimited reasoning, removes it from main text, and emits reasoning separately.
Key points
- It separates delimited reasoning sections from the model's generated text.
- Extracted reasoning is made available independently from the main response text.
- The middleware supports both streamed and complete model responses.
- Reasoning and its delimiters are excluded from the main text result.
- Streaming output can contain distinct reasoning and text content parts.
- The optional separator controls how remaining text sections are joined.
Common mistakes
- Use an object with literal opening and closing delimiters when the model does not emit XML-style tags.
- Enable startWithReasoning when reasoning begins immediately without its opening delimiter.
Example: Extract tagged reasoning
import { generateText, wrapLanguageModel, extractReasoningMiddleware } from 'ai';
import { yourModel } from './model';
const model = wrapLanguageModel({
model: yourModel,
middleware: extractReasoningMiddleware({ tagName: 'think' }),
});
const { reasoning, text } = await generateText({
model,
prompt: 'Explain quantum entanglement.',
});
console.log({ reasoning, text });Wraps a language model so content inside <think> tags is returned separately as reasoning while the remaining response is available as text.
When to use it
Handle complete and streaming output
Process tagged reasoning while consuming generateText or streamText results.
Support custom delimiters
Adapt the middleware to providers that use non-XML markers.
Related
wrapLanguageModelgenerateTextstreamText