Implementation:Ollama Ollama XModels GLM4 Parser

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

Overview

State-machine parser that extracts thinking content, regular content, and tool calls from GLM4-MoE-Lite model output.

Description

Implements a seven-state parser tracking the parsing context: looking for thinking open tag, collecting thinking, collecting content, and collecting tool content. When thinking is enabled, the parser starts in CollectingThinking state since GLM4's prompt ends with <think>. The eat method consumes buffer content based on current state, splitting at XML-like tags (<think>, </think>, <tool_call>, </tool_call>). Tool calls are parsed from XML format into structured api.ToolCall objects.

Usage

Used when serving GLM4-MoE-Lite models to parse the streaming text output into structured thinking, content, and tool call events.

Code Reference

Source Location

• Repository: Ollama
• File: x/models/glm4_moe_lite/parser.go
• Lines: 1-479

Signature

type Parser struct {
    state  parserState
    buffer strings.Builder
    tools  []api.Tool
}

func (p *Parser) HasToolSupport() bool
func (p *Parser) HasThinkingSupport() bool
func (p *Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool
func (p *Parser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error)

Import

import "github.com/ollama/ollama/x/models/glm4_moe_lite"

I/O Contract

Inputs

Name	Type	Required	Description
s	string	Yes	New output text chunk from the model
done	bool	Yes	Whether this is the final chunk

Outputs

Name	Type	Description
content	string	Regular content extracted from the output
thinking	string	Thinking/reasoning content extracted
calls	[]api.ToolCall	Parsed tool calls
err	error	Parsing error if any

Usage Examples

parser := &Parser{}
parser.Init(tools, nil, nil)

// Process streaming output
content, thinking, calls, err := parser.Add("<think>Let me analyze...", false)
content, thinking, calls, err = parser.Add("</think>The answer is 42.", true)

Related Pages

• Principle:Ollama_Ollama_ModelArchitecture