feat(gemma4): add thinking support

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

Author: mudler

core/http/react-ui/src/hooks/useChat.js

@@ -2,9 +2,9 @@ import { useState, useCallback, useRef, useEffect } from 'react'
 import { API_CONFIG } from '../utils/config'
 import { apiUrl } from '../utils/basePath'
 
-const thinkingTagRegex = /<thinking>([\s\S]*?)<\/thinking>|<think>([\s\S]*?)<\/think>/g
-const openThinkTagRegex = /<thinking>|<think>/
-const closeThinkTagRegex = /<\/thinking>|<\/think>/
+const thinkingTagRegex = /<thinking>([\s\S]*?)<\/thinking>|<think>([\s\S]*?)<\/think>|<\|channel>thought([\s\S]*?)<channel\|>/g
+const openThinkTagRegex = /<thinking>|<think>|<\|channel>thought/
+const closeThinkTagRegex = /<\/thinking>|<\/think>|<channel\|>/
 
 async function extractHttpError(response) {
   let errorMsg = `HTTP ${response.status}`
@@ -23,7 +23,7 @@ function extractThinking(text) {
   thinkingTagRegex.lastIndex = 0
   while ((match = thinkingTagRegex.exec(text)) !== null) {
     regularContent += text.slice(lastIdx, match.index)
-    thinkingContent += match[1] || match[2] || ''
+    thinkingContent += match[1] || match[2] || match[3] || ''
     lastIdx = match.index + match[0].length
   }
   regularContent += text.slice(lastIdx)
@@ -578,9 +578,9 @@ export function useChat(initialModel = '') {
                     }
 
                     if (insideThinkTag) {
-                      const lastOpen = Math.max(rawContent.lastIndexOf('<thinking>'), rawContent.lastIndexOf('<think>'))
+                      const lastOpen = Math.max(rawContent.lastIndexOf('<thinking>'), rawContent.lastIndexOf('<think>'), rawContent.lastIndexOf('<|channel>thought'))
                       if (lastOpen >= 0) {
-                        const partial = rawContent.slice(lastOpen).replace(/<thinking>|<think>/, '')
+                        const partial = rawContent.slice(lastOpen).replace(/<thinking>|<think>|<\|channel>thought/, '')
                         setStreamingReasoning(partial)
                         const beforeThink = rawContent.slice(0, lastOpen)
                         const { regularContent: contentBeforeThink } = extractThinking(beforeThink)

pkg/reasoning/reasoning.go

@@ -15,6 +15,7 @@ import (
 // - <|inner_prefix|>        (Apertus models)
 // - <seed:think>            (Seed models)
 // - <think>    (DeepSeek, Granite, ExaOne models)
+// - <|channel>thought        (Gemma 4 models)
 // - <|think|>               (Solar Open models)
 // - <thinking>              (General thinking tag)
 // - [THINK]                 (Magistral models)
@@ -28,6 +29,7 @@ func DetectThinkingStartToken(prompt string, config *Config) string {
 		"<seed:think>",       // Seed models
 		"<think>",            // DeepSeek, Granite, ExaOne models
 		"<|think|>",          // Solar Open models
+		"<|channel>thought",  // Gemma 4 models
 		"<thinking>",         // General thinking tag
 		"[THINK]",            // Magistral models
 	}
@@ -146,6 +148,7 @@ func ExtractReasoning(content string, config *Config) (reasoning string, cleaned
 		{"<seed:think>", "</seed:think>"},                     // Seed models
 		{"<think>", "</think>"},                               // DeepSeek, Granite, ExaOne models
 		{"<|think|>", "<|end|><|begin|>assistant<|content|>"}, // Solar Open models (complex end)
+		{"<|channel>thought", "<channel|>"},                    // Gemma 4 models
 		{"<thinking>", "</thinking>"},                         // General thinking tag
 		{"[THINK]", "[/THINK]"},                               // Magistral models
 	}

pkg/reasoning/reasoning_test.go

@@ -317,6 +317,29 @@ var _ = Describe("ExtractReasoning", func() {
 			Expect(cleaned).To(Equal("Before "))
 		})
 	})
+
+	Context("when content has <|channel>thought tags (Gemma 4)", func() {
+		It("should extract reasoning from channel thought block", func() {
+			content := "<|channel>thought\nThis is my reasoning\n<channel|>Hello! How can I help?"
+			reasoning, cleaned := ExtractReasoning(content, nil)
+			Expect(reasoning).To(Equal("This is my reasoning"))
+			Expect(cleaned).To(Equal("Hello! How can I help?"))
+		})
+
+		It("should handle unclosed channel thought block", func() {
+			content := "<|channel>thought\nIncomplete reasoning"
+			reasoning, cleaned := ExtractReasoning(content, nil)
+			Expect(reasoning).To(Equal("Incomplete reasoning"))
+			Expect(cleaned).To(Equal(""))
+		})
+
+		It("should handle content before and after channel thought block", func() {
+			content := "Before <|channel>thought\nGemma 4 reasoning\n<channel|> After"
+			reasoning, cleaned := ExtractReasoning(content, nil)
+			Expect(reasoning).To(Equal("Gemma 4 reasoning"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+	})
 })
 
 var _ = Describe("DetectThinkingStartToken", func() {
@@ -339,6 +362,12 @@ var _ = Describe("DetectThinkingStartToken", func() {
 			Expect(token).To(Equal("<thinking>"))
 		})
 
+		It("should detect <|channel>thought at the end (Gemma 4)", func() {
+			prompt := "Prompt text <|channel>thought"
+			token := DetectThinkingStartToken(prompt, nil)
+			Expect(token).To(Equal("<|channel>thought"))
+		})
+
 		It("should detect <|inner_prefix|> at the end", func() {
 			prompt := "Prompt <|inner_prefix|>"
 			token := DetectThinkingStartToken(prompt, nil)
@@ -817,6 +846,14 @@ var _ = Describe("ExtractReasoningWithConfig", func() {
 			Expect(cleaned).To(Equal("Text  More"))
 		})
 
+		It("should strip reasoning from Gemma 4 channel tags when StripReasoningOnly is true", func() {
+			content := "<|channel>thought\nGemma 4 reasoning\n<channel|>Response text"
+			config := Config{StripReasoningOnly: boolPtr(true)}
+			reasoning, cleaned := ExtractReasoningWithConfig(content, "<|channel>thought", config)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal("Response text"))
+		})
+
 		It("should strip reasoning with multiline content when StripReasoningOnly is true", func() {
 			content := "Start <thinking>Line 1\nLine 2\nLine 3</thinking> End"
 			config := Config{StripReasoningOnly: boolPtr(true)}