Implementation:Ollama Ollama Thinking Parser

Knowledge Sources	Ollama
Domains	Thinking, Output Parsing
Last Updated	2025-02-15 00:00 GMT

Overview

State-machine parser that separates LLM output into thinking content (reasoning traces) and regular content based on configurable opening/closing tags.

Description

The Parser tracks five states: looking for opening tag, eating whitespace after opening, collecting thinking content, eating whitespace after closing tag, and done thinking. AddContent accepts streaming text chunks and returns thinking and remaining content separately. The parser handles partial tag matches by buffering ambiguous content until it can be disambiguated. If no opening tag is found (model skips thinking), all content is returned as regular content. The overlap helper detects partial tag matches at chunk boundaries to avoid splitting tags across calls.

Usage

Used in the streaming response pipeline to separate thinking/reasoning content from final answers, enabling the client to display them differently.

Code Reference

Source Location

• Repository: Ollama
• File: thinking/parser.go
• Lines: 1-173

Signature

type Parser struct {
    state      thinkingState
    OpeningTag string
    ClosingTag string
    acc        strings.Builder
}

func (s *Parser) AddContent(content string) (thinking string, remaining string)

Import

import "github.com/ollama/ollama/thinking"

I/O Contract

Inputs

Name	Type	Required	Description
content	string	Yes	Streaming text chunk from the model

Outputs

Name	Type	Description
thinking	string	Thinking/reasoning content extracted from this chunk
remaining	string	Regular content to display to the user

Usage Examples

parser := &thinking.Parser{
    OpeningTag: "<think>",
    ClosingTag: "</think>",
}

// Process streaming output
t1, r1 := parser.AddContent("<think>Let me")
// t1="", r1="" (buffering partial tag)

t2, r2 := parser.AddContent(" analyze this...")
// t2="Let me analyze this...", r2=""

t3, r3 := parser.AddContent("</think>The answer is 42.")
// t3="", r3="The answer is 42."

Related Pages

• Principle:Ollama_Ollama_ThinkingSupport