Additional thinking tags

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

core/config/model_config.go

@@ -10,6 +10,7 @@ import (
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/pkg/downloader"
 	"github.com/mudler/LocalAI/pkg/functions"
+	"github.com/mudler/LocalAI/pkg/reasoning"
 	"github.com/mudler/cogito"
 	"gopkg.in/yaml.v3"
 )
@@ -51,6 +52,7 @@ type ModelConfig struct {
 	ResponseFormatMap                          map[string]interface{} `yaml:"-" json:"-"`
 
 	FunctionsConfig functions.FunctionsConfig `yaml:"function,omitempty" json:"function,omitempty"`
+	ReasoningConfig reasoning.ReasoningConfig `yaml:"reasoning,omitempty" json:"reasoning,omitempty"`
 
 	FeatureFlag FeatureFlag `yaml:"feature_flags,omitempty" json:"feature_flags,omitempty"` // Feature Flag registry. We move fast, and features may break on a per model/backend basis. Registry for (usually temporary) flags that indicate aborting something early.
 	// LLM configs (GPT4ALL, Llama.cpp, ...)

core/http/endpoints/openai/chat.go

@@ -13,6 +13,7 @@ import (
 	"github.com/mudler/LocalAI/core/http/middleware"
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/pkg/functions"
+	"github.com/mudler/LocalAI/pkg/reasoning"
 
 	"github.com/mudler/LocalAI/core/templates"
 	"github.com/mudler/LocalAI/pkg/model"
@@ -43,10 +44,19 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 		lastEmittedReasoning := ""
 		lastEmittedCleanedContent := ""
 
+		// Configure reasoning extraction options
+		// Auto-detect if prompt ends with thinking tag
+		// or use explicit config setting
+		thinkingForcedOpen := config.ReasoningConfig.ThinkingForcedOpen || reasoning.DetectThinkingForcedOpen(s)
+
 		_, _, err := ComputeChoices(req, s, config, cl, startupOptions, loader, func(s string, c *[]schema.Choice) {}, func(s string, tokenUsage backend.TokenUsage) bool {
 			accumulatedContent += s
 			// Extract reasoning from accumulated content
-			currentReasoning, cleanedContent := functions.ExtractReasoning(accumulatedContent)
+			opts := []reasoning.Option{}
+			if thinkingForcedOpen {
+				opts = append(opts, reasoning.WithThinkingForcedOpen())
+			}
+			currentReasoning, cleanedContent := reasoning.Extract(accumulatedContent, opts...)
 
 			// Calculate new reasoning delta (what we haven't emitted yet)
 			var reasoningDelta *string
@@ -230,7 +240,13 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 			return err
 		}
 		// Extract reasoning before processing tool calls
-		reasoning, cleanedResult := functions.ExtractReasoning(result)
+		// Auto-detect if prompt ends with thinking tag or use explicit config
+		toolsThinkingForcedOpen := config.ReasoningConfig.ThinkingForcedOpen || reasoning.DetectThinkingForcedOpen(prompt)
+		opts := []reasoning.Option{}
+		if toolsThinkingForcedOpen {
+			opts = append(opts, reasoning.WithThinkingForcedOpen())
+		}
+		extractedReasoning, cleanedResult := reasoning.Extract(result, opts...)
 		result = cleanedResult
 
 		textContentToReturn = functions.ParseTextContent(result, config.FunctionsConfig)
@@ -266,8 +282,8 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 			}
 
 			var deltaReasoning *string
-			if reasoning != "" {
-				deltaReasoning = &reasoning
+			if extractedReasoning != "" {
+				deltaReasoning = &extractedReasoning
 			}
 			delta := &schema.Message{Content: &result}
 			if deltaReasoning != nil {
@@ -618,17 +634,24 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 		// no streaming mode
 		default:
 
+			// Auto-detect if prompt ends with thinking tag for non-streaming mode
+			nonStreamThinkingForcedOpen := config.ReasoningConfig.ThinkingForcedOpen || reasoning.DetectThinkingForcedOpen(predInput)
+
 			tokenCallback := func(s string, c *[]schema.Choice) {
 				// Extract reasoning from the response
-				reasoning, cleanedS := functions.ExtractReasoning(s)
-				s = cleanedS
+				var extractedReasoning string
+				opts := []reasoning.Option{}
+				if nonStreamThinkingForcedOpen {
+					opts = append(opts, reasoning.WithThinkingForcedOpen())
+				}
+				extractedReasoning, s = reasoning.Extract(s, opts...)
 
 				if !shouldUseFn {
 					// no function is called, just reply and use stop as finish reason
 					stopReason := FinishReasonStop
 					message := &schema.Message{Role: "assistant", Content: &s}
-					if reasoning != "" {
-						message.Reasoning = &reasoning
+					if extractedReasoning != "" {
+						message.Reasoning = &extractedReasoning
 					}
 					*c = append(*c, schema.Choice{FinishReason: &stopReason, Index: 0, Message: message})
 					return
@@ -650,8 +673,8 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 
 					stopReason := FinishReasonStop
 					message := &schema.Message{Role: "assistant", Content: &result}
-					if reasoning != "" {
-						message.Reasoning = &reasoning
+					if extractedReasoning != "" {
+						message.Reasoning = &extractedReasoning
 					}
 					*c = append(*c, schema.Choice{
 						FinishReason: &stopReason,
@@ -664,8 +687,8 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 							Role: "assistant",
 						},
 					}
-					if reasoning != "" {
-						toolChoice.Message.Reasoning = &reasoning
+					if extractedReasoning != "" {
+						toolChoice.Message.Reasoning = &extractedReasoning
 					}
 
 					for _, ss := range results {
@@ -695,8 +718,8 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
 									"arguments": args,
 								},
 							}
-							if reasoning != "" {
-								message.Reasoning = &reasoning
+							if extractedReasoning != "" {
+								message.Reasoning = &extractedReasoning
 							}
 							*c = append(*c, schema.Choice{
 								FinishReason: &functionCallReason,

pkg/functions/reasoning.go

@@ -1,114 +0,0 @@
-package functions
-
-import (
-	"strings"
-)
-
-// ExtractReasoning extracts reasoning content from thinking tags and returns
-// both the extracted reasoning and the cleaned content (with tags removed).
-// It handles <thinking>...</thinking> and <think>...</think> tags.
-// Multiple reasoning blocks are concatenated with newlines.
-func ExtractReasoning(content string) (reasoning string, cleanedContent string) {
-	if content == "" {
-		return "", content
-	}
-
-	var reasoningParts []string
-	var cleanedParts []string
-	remaining := content
-
-	// Define tag pairs to look for
-	tagPairs := []struct {
-		start string
-		end   string
-	}{
-		{"<thinking>", "</thinking>"},
-		{"<think>", "</think>"},
-	}
-
-	// Track the last position we've processed
-	lastPos := 0
-
-	for {
-		// Find the earliest tag start
-		earliestStart := -1
-		earliestEnd := -1
-		isUnclosed := false
-		var matchedTag struct {
-			start string
-			end   string
-		}
-
-		for _, tagPair := range tagPairs {
-			startIdx := strings.Index(remaining[lastPos:], tagPair.start)
-			if startIdx == -1 {
-				continue
-			}
-			startIdx += lastPos
-
-			// Find the corresponding end tag
-			endIdx := strings.Index(remaining[startIdx+len(tagPair.start):], tagPair.end)
-			if endIdx == -1 {
-				// Unclosed tag - extract what we have
-				if earliestStart == -1 || startIdx < earliestStart {
-					earliestStart = startIdx
-					earliestEnd = len(remaining)
-					isUnclosed = true
-					matchedTag = tagPair
-				}
-				continue
-			}
-			endIdx += startIdx + len(tagPair.start)
-
-			// Found a complete tag pair
-			if earliestStart == -1 || startIdx < earliestStart {
-				earliestStart = startIdx
-				earliestEnd = endIdx + len(tagPair.end)
-				isUnclosed = false
-				matchedTag = tagPair
-			}
-		}
-
-		if earliestStart == -1 {
-			// No more tags found, add remaining content
-			if lastPos < len(remaining) {
-				cleanedParts = append(cleanedParts, remaining[lastPos:])
-			}
-			break
-		}
-
-		// Add content before the tag
-		if earliestStart > lastPos {
-			cleanedParts = append(cleanedParts, remaining[lastPos:earliestStart])
-		}
-
-		// Extract reasoning content
-		reasoningStart := earliestStart + len(matchedTag.start)
-		// For unclosed tags, earliestEnd is already at the end of the string
-		// For closed tags, earliestEnd points to after the closing tag, so we subtract the end tag length
-		var reasoningEnd int
-		if isUnclosed {
-			// Unclosed tag - extract everything to the end
-			reasoningEnd = len(remaining)
-		} else {
-			// Closed tag - exclude the end tag
-			reasoningEnd = earliestEnd - len(matchedTag.end)
-		}
-		if reasoningEnd > reasoningStart {
-			reasoningContent := strings.TrimSpace(remaining[reasoningStart:reasoningEnd])
-			if reasoningContent != "" {
-				reasoningParts = append(reasoningParts, reasoningContent)
-			}
-		}
-
-		// Move past this tag
-		lastPos = earliestEnd
-	}
-
-	// Combine reasoning parts
-	reasoning = strings.Join(reasoningParts, "\n\n")
-	// Combine cleaned content parts
-	cleanedContent = strings.Join(cleanedParts, "")
-
-	return reasoning, cleanedContent
-}

pkg/functions/reasoning_test.go

@@ -1,261 +0,0 @@
-package functions_test
-
-import (
-	"strings"
-
-	. "github.com/mudler/LocalAI/pkg/functions"
-	. "github.com/onsi/ginkgo/v2"
-	. "github.com/onsi/gomega"
-)
-
-var _ = Describe("ExtractReasoning", func() {
-	Context("when content has no reasoning tags", func() {
-		It("should return empty reasoning and original content", func() {
-			content := "This is regular content without any tags."
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(BeEmpty())
-			Expect(cleaned).To(Equal(content))
-		})
-
-		It("should handle empty string", func() {
-			content := ""
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(BeEmpty())
-			Expect(cleaned).To(BeEmpty())
-		})
-
-		It("should handle content with only whitespace", func() {
-			content := "   \n\t  "
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(BeEmpty())
-			Expect(cleaned).To(Equal(content))
-		})
-	})
-
-	Context("when content has <thinking> tags", func() {
-		It("should extract reasoning from single thinking block", func() {
-			content := "Some text <thinking>This is my reasoning</thinking> More text"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("This is my reasoning"))
-			Expect(cleaned).To(Equal("Some text  More text"))
-		})
-
-		It("should extract reasoning and preserve surrounding content", func() {
-			content := "Before <thinking>Reasoning here</thinking> After"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Reasoning here"))
-			Expect(cleaned).To(Equal("Before  After"))
-		})
-
-		It("should handle thinking block at the start", func() {
-			content := "<thinking>Start reasoning</thinking> Regular content"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Start reasoning"))
-			Expect(cleaned).To(Equal(" Regular content"))
-		})
-
-		It("should handle thinking block at the end", func() {
-			content := "Regular content <thinking>End reasoning</thinking>"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("End reasoning"))
-			Expect(cleaned).To(Equal("Regular content "))
-		})
-
-		It("should handle only thinking block", func() {
-			content := "<thinking>Only reasoning</thinking>"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Only reasoning"))
-			Expect(cleaned).To(BeEmpty())
-		})
-
-		It("should trim whitespace from reasoning content", func() {
-			content := "Text <thinking>  \n  Reasoning with spaces  \n  </thinking> More"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Reasoning with spaces"))
-			Expect(cleaned).To(Equal("Text  More"))
-		})
-	})
-
-	Context("when content has <think> tags", func() {
-		It("should extract reasoning from redacted_reasoning block", func() {
-			content := "Text <think>Redacted reasoning</think> More"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Redacted reasoning"))
-			Expect(cleaned).To(Equal("Text  More"))
-		})
-
-		It("should handle redacted_reasoning with multiline content", func() {
-			content := "Before <think>Line 1\nLine 2\nLine 3</think> After"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Line 1\nLine 2\nLine 3"))
-			Expect(cleaned).To(Equal("Before  After"))
-		})
-
-		It("should handle redacted_reasoning with complex content", func() {
-			content := "Start <think>Complex reasoning\nwith\nmultiple\nlines</think> End"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Complex reasoning\nwith\nmultiple\nlines"))
-			Expect(cleaned).To(Equal("Start  End"))
-		})
-	})
-
-	Context("when content has multiple reasoning blocks", func() {
-		It("should concatenate multiple thinking blocks with newlines", func() {
-			content := "Text <thinking>First</thinking> Middle <thinking>Second</thinking> End"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("First\n\nSecond"))
-			Expect(cleaned).To(Equal("Text  Middle  End"))
-		})
-
-		It("should handle multiple different tag types", func() {
-			content := "A <thinking>One</thinking> B <think>Two</think> C <think>Three</think> D"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(ContainSubstring("One"))
-			Expect(reasoning).To(ContainSubstring("Two"))
-			Expect(reasoning).To(ContainSubstring("Three"))
-			Expect(cleaned).To(Equal("A  B  C  D"))
-		})
-
-		It("should handle nested tags correctly (extracts first match)", func() {
-			content := "Text <thinking>Outer <think>Inner</think></thinking> More"
-			reasoning, cleaned := ExtractReasoning(content)
-			// Should extract the outer thinking block
-			Expect(reasoning).To(ContainSubstring("Outer"))
-			Expect(reasoning).To(ContainSubstring("Inner"))
-			Expect(cleaned).To(Equal("Text  More"))
-		})
-	})
-
-	Context("when content has unclosed reasoning tags", func() {
-		It("should extract unclosed thinking block", func() {
-			content := "Text <thinking>Unclosed reasoning"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Unclosed reasoning"))
-			Expect(cleaned).To(Equal("Text "))
-		})
-
-		It("should extract unclosed think block", func() {
-			content := "Before <think>Incomplete"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Incomplete"))
-			Expect(cleaned).To(Equal("Before "))
-		})
-
-		It("should extract unclosed redacted_reasoning block", func() {
-			content := "Start <think>Partial reasoning content"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Partial reasoning content"))
-			Expect(cleaned).To(Equal("Start "))
-		})
-
-		It("should handle unclosed tag at the end", func() {
-			content := "Regular content <thinking>Unclosed at end"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Unclosed at end"))
-			Expect(cleaned).To(Equal("Regular content "))
-		})
-	})
-
-	Context("when content has empty reasoning blocks", func() {
-		It("should ignore empty thinking block", func() {
-			content := "Text <thinking></thinking> More"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(BeEmpty())
-			Expect(cleaned).To(Equal("Text  More"))
-		})
-
-		It("should ignore thinking block with only whitespace", func() {
-			content := "Text <thinking>   \n\t  </thinking> More"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(BeEmpty())
-			Expect(cleaned).To(Equal("Text  More"))
-		})
-	})
-
-	Context("when content has reasoning tags with special characters", func() {
-		It("should handle reasoning with newlines", func() {
-			content := "Before <thinking>Line 1\nLine 2\nLine 3</thinking> After"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Line 1\nLine 2\nLine 3"))
-			Expect(cleaned).To(Equal("Before  After"))
-		})
-
-		It("should handle reasoning with code blocks", func() {
-			content := "Text <thinking>Reasoning with ```code``` blocks</thinking> More"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Reasoning with ```code``` blocks"))
-			Expect(cleaned).To(Equal("Text  More"))
-		})
-
-		It("should handle reasoning with JSON", func() {
-			content := "Before <think>{\"key\": \"value\"}</think> After"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("{\"key\": \"value\"}"))
-			Expect(cleaned).To(Equal("Before  After"))
-		})
-
-		It("should handle reasoning with HTML-like content", func() {
-			content := "Text <thinking>Reasoning with <tags> inside</thinking> More"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Reasoning with <tags> inside"))
-			Expect(cleaned).To(Equal("Text  More"))
-		})
-	})
-
-	Context("when content has reasoning mixed with regular content", func() {
-		It("should preserve content order correctly", func() {
-			content := "Start <thinking>Reasoning</thinking> Middle <think>More reasoning</think> End"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(ContainSubstring("Reasoning"))
-			Expect(reasoning).To(ContainSubstring("More reasoning"))
-			Expect(cleaned).To(Equal("Start  Middle  End"))
-		})
-
-		It("should handle reasoning in the middle of a sentence", func() {
-			content := "This is a <thinking>reasoning</thinking> sentence."
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("reasoning"))
-			Expect(cleaned).To(Equal("This is a  sentence."))
-		})
-	})
-
-	Context("edge cases", func() {
-		It("should handle content with only opening tag", func() {
-			content := "<thinking>"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(BeEmpty())
-			Expect(cleaned).To(Equal(""))
-		})
-
-		It("should handle content with only closing tag", func() {
-			content := "</thinking>"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(BeEmpty())
-			Expect(cleaned).To(Equal("</thinking>"))
-		})
-
-		It("should handle mismatched tags", func() {
-			content := "<thinking>Content</think>"
-			reasoning, cleaned := ExtractReasoning(content)
-			// Should extract unclosed thinking block
-			Expect(reasoning).To(ContainSubstring("Content"))
-			Expect(cleaned).To(Equal(""))
-		})
-
-		It("should handle very long reasoning content", func() {
-			longReasoning := strings.Repeat("This is reasoning content. ", 100)
-			content := "Text <thinking>" + longReasoning + "</thinking> More"
-			reasoning, cleaned := ExtractReasoning(content)
-			// TrimSpace is applied, so we need to account for that
-			Expect(reasoning).To(Equal(strings.TrimSpace(longReasoning)))
-			Expect(cleaned).To(Equal("Text  More"))
-		})
-
-		It("should handle reasoning with unicode characters", func() {
-			content := "Text <thinking>Reasoning with 中文 and emoji 🧠</thinking> More"
-			reasoning, cleaned := ExtractReasoning(content)
-			Expect(reasoning).To(Equal("Reasoning with 中文 and emoji 🧠"))
-			Expect(cleaned).To(Equal("Text  More"))
-		})
-	})
-})

pkg/reasoning/config.go

@@ -0,0 +1,8 @@
+package reasoning
+
+type ReasoningConfig struct {
+	// ThinkingForcedOpen indicates that the model outputs reasoning without an opening tag.
+	// When true, all content from the start is treated as reasoning until a closing tag is found.
+	// This is useful for models like GLM-4 that output reasoning without <think> but end with </think>.
+	ThinkingForcedOpen bool `yaml:"thinking_forced_open,omitempty" json:"thinking_forced_open,omitempty"`
+}

pkg/reasoning/options.go

@@ -0,0 +1,18 @@
+package reasoning
+
+// options holds the configuration for reasoning extraction
+type options struct {
+	thinkingForcedOpen bool
+}
+
+// Option is a functional option for configuring reasoning extraction
+type Option func(*options)
+
+// WithThinkingForcedOpen configures the extractor to treat all content from the start
+// as reasoning until a closing tag is found. This is useful for models like GLM-4
+// that output reasoning without <think> but end with </think>.
+func WithThinkingForcedOpen() Option {
+	return func(o *options) {
+		o.thinkingForcedOpen = true
+	}
+}

pkg/reasoning/reasoning.go

@@ -0,0 +1,256 @@
+package reasoning
+
+import (
+	"strings"
+)
+
+// Common thinking/reasoning opening tags used by various models.
+// These match the tags detected by llama.cpp in common/chat.cpp
+var thinkingOpenTags = []string{
+	// DeepSeek R1, V3.1, Nemotron V2, MiniMax M2, Hermes 2 Pro, Granite, Exaone MOE
+	"<think>\n",
+	"<think>",
+	// Generic thinking tags
+	"<thinking>\n",
+	"<thinking>",
+	// Apertus
+	"<|inner_prefix|>",
+	// Command R7B
+	"<|START_THINKING|>",
+	// Seed
+	"<seed:think>",
+	// Magistral (not in llama.cpp but common)
+	"[THINK]\n",
+	"[THINK]",
+}
+
+// DetectThinkingForcedOpen checks if a prompt ends with a thinking opening tag.
+// This is used to automatically detect when the model template has already added
+// the opening thinking tag, meaning the model will output reasoning content directly.
+// Returns true if the prompt ends with a known thinking opening tag.
+func DetectThinkingForcedOpen(prompt string) bool {
+	for _, tag := range thinkingOpenTags {
+		if strings.HasSuffix(prompt, tag) {
+			return true
+		}
+	}
+	return false
+}
+
+// Extract extracts reasoning content from thinking tags and returns
+// both the extracted reasoning and the cleaned content (with tags removed).
+// It handles <thinking>...</thinking> and <think>...</think> tags.
+// Multiple reasoning blocks are concatenated with newlines.
+// It also handles the case where only a closing tag is present (no opening tag),
+// in which case everything before the closing tag is treated as reasoning.
+//
+// Use WithThinkingForcedOpen() option when all content from the start should be
+// treated as reasoning until a closing tag is found.
+func Extract(content string, opts ...Option) (reasoning string, cleanedContent string) {
+	if content == "" {
+		return "", content
+	}
+
+	cfg := &options{}
+	for _, opt := range opts {
+		opt(cfg)
+	}
+
+	if cfg.thinkingForcedOpen {
+		return extractForcedOpen(content)
+	}
+
+	return extractFromTags(content)
+}
+
+// extractForcedOpen handles the case where reasoning starts without an opening tag.
+// All content from the start is treated as reasoning until a closing tag is found.
+func extractForcedOpen(content string) (reasoning string, cleanedContent string) {
+	// Look for the earliest closing tag
+	// These match the closing tags used by llama.cpp for various models
+	closingTags := []string{
+		"</thinking>",
+		"</think>",
+		"<|END_THINKING|>", // Command R7B
+		"<|inner_suffix|>", // Apertus
+		"</seed:think>",    // Seed
+		"[/THINK]",         // Magistral
+	}
+
+	earliestCloseIdx := -1
+	var matchedCloseTag string
+
+	for _, closeTag := range closingTags {
+		idx := strings.Index(content, closeTag)
+		if idx != -1 && (earliestCloseIdx == -1 || idx < earliestCloseIdx) {
+			earliestCloseIdx = idx
+			matchedCloseTag = closeTag
+		}
+	}
+
+	if earliestCloseIdx == -1 {
+		// No closing tag found - all content is reasoning (still streaming)
+		return strings.TrimSpace(content), ""
+	}
+
+	// Found closing tag - everything before is reasoning, everything after is content
+	reasoning = strings.TrimSpace(content[:earliestCloseIdx])
+	cleanedContent = content[earliestCloseIdx+len(matchedCloseTag):]
+
+	// Continue processing the rest for any additional reasoning blocks
+	if cleanedContent != "" {
+		additionalReasoning, finalContent := extractFromTags(cleanedContent)
+		if additionalReasoning != "" {
+			if reasoning != "" {
+				reasoning = reasoning + "\n\n" + additionalReasoning
+			} else {
+				reasoning = additionalReasoning
+			}
+		}
+		cleanedContent = finalContent
+	}
+
+	return reasoning, cleanedContent
+}
+
+// extractFromTags extracts reasoning content from thinking tags.
+// This is the core implementation that handles standard tag-based extraction.
+func extractFromTags(content string) (reasoning string, cleanedContent string) {
+	if content == "" {
+		return "", content
+	}
+
+	var reasoningParts []string
+	var cleanedParts []string
+	remaining := content
+
+	// Define tag pairs to look for
+	// These match the tags used by llama.cpp for various models
+	tagPairs := []struct {
+		start string
+		end   string
+	}{
+		{"<thinking>", "</thinking>"},
+		{"<think>", "</think>"},
+		{"<|START_THINKING|>", "<|END_THINKING|>"}, // Command R7B
+		{"<|inner_prefix|>", "<|inner_suffix|>"},   // Apertus
+		{"<seed:think>", "</seed:think>"},          // Seed
+		{"[THINK]", "[/THINK]"},                    // Magistral
+	}
+
+	// Track the last position we've processed
+	lastPos := 0
+
+	for {
+		// Find the earliest tag start
+		earliestStart := -1
+		earliestEnd := -1
+		isUnclosed := false
+		isClosingOnly := false
+		var matchedTag struct {
+			start string
+			end   string
+		}
+
+		for _, tagPair := range tagPairs {
+			startIdx := strings.Index(remaining[lastPos:], tagPair.start)
+			endIdx := strings.Index(remaining[lastPos:], tagPair.end)
+
+			// Check for closing-only tag (closing tag appears before or without opening tag)
+			if endIdx != -1 && (startIdx == -1 || endIdx < startIdx) {
+				// Found a closing tag without a preceding opening tag
+				closingTagPos := endIdx + lastPos
+				if earliestStart == -1 || closingTagPos < earliestStart || (isClosingOnly && closingTagPos < earliestEnd) {
+					earliestStart = lastPos
+					earliestEnd = closingTagPos + len(tagPair.end)
+					isClosingOnly = true
+					isUnclosed = false
+					matchedTag = tagPair
+				}
+				continue
+			}
+
+			if startIdx == -1 {
+				continue
+			}
+			startIdx += lastPos
+
+			// Find the corresponding end tag after the start tag
+			endIdxAfterStart := strings.Index(remaining[startIdx+len(tagPair.start):], tagPair.end)
+			if endIdxAfterStart == -1 {
+				// Unclosed tag - extract what we have
+				if earliestStart == -1 || startIdx < earliestStart {
+					earliestStart = startIdx
+					earliestEnd = len(remaining)
+					isUnclosed = true
+					isClosingOnly = false
+					matchedTag = tagPair
+				}
+				continue
+			}
+			endIdxAfterStart += startIdx + len(tagPair.start)
+
+			// Found a complete tag pair
+			if earliestStart == -1 || startIdx < earliestStart {
+				earliestStart = startIdx
+				earliestEnd = endIdxAfterStart + len(tagPair.end)
+				isUnclosed = false
+				isClosingOnly = false
+				matchedTag = tagPair
+			}
+		}
+
+		if earliestStart == -1 {
+			// No more tags found, add remaining content
+			if lastPos < len(remaining) {
+				cleanedParts = append(cleanedParts, remaining[lastPos:])
+			}
+			break
+		}
+
+		if isClosingOnly {
+			// Closing tag without opening tag - content before closing tag is reasoning
+			reasoningContent := strings.TrimSpace(remaining[lastPos : earliestEnd-len(matchedTag.end)])
+			if reasoningContent != "" {
+				reasoningParts = append(reasoningParts, reasoningContent)
+			}
+			// Move past the closing tag
+			lastPos = earliestEnd
+			continue
+		}
+
+		// Add content before the tag
+		if earliestStart > lastPos {
+			cleanedParts = append(cleanedParts, remaining[lastPos:earliestStart])
+		}
+
+		// Extract reasoning content
+		reasoningStart := earliestStart + len(matchedTag.start)
+		// For unclosed tags, earliestEnd is already at the end of the string
+		// For closed tags, earliestEnd points to after the closing tag, so we subtract the end tag length
+		var reasoningEnd int
+		if isUnclosed {
+			// Unclosed tag - extract everything to the end
+			reasoningEnd = len(remaining)
+		} else {
+			// Closed tag - exclude the end tag
+			reasoningEnd = earliestEnd - len(matchedTag.end)
+		}
+		if reasoningEnd > reasoningStart {
+			reasoningContent := strings.TrimSpace(remaining[reasoningStart:reasoningEnd])
+			if reasoningContent != "" {
+				reasoningParts = append(reasoningParts, reasoningContent)
+			}
+		}
+
+		// Move past this tag
+		lastPos = earliestEnd
+	}
+
+	// Combine reasoning parts
+	reasoning = strings.Join(reasoningParts, "\n\n")
+	// Combine cleaned content parts
+	cleanedContent = strings.Join(cleanedParts, "")
+
+	return reasoning, cleanedContent
+}

pkg/reasoning/reasoning_suite_test.go

@@ -0,0 +1,13 @@
+package reasoning_test
+
+import (
+	"testing"
+
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+func TestReasoning(t *testing.T) {
+	RegisterFailHandler(Fail)
+	RunSpecs(t, "Reasoning Suite")
+}

pkg/reasoning/reasoning_test.go

@@ -0,0 +1,499 @@
+package reasoning_test
+
+import (
+	"strings"
+
+	. "github.com/mudler/LocalAI/pkg/reasoning"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("DetectThinkingForcedOpen", func() {
+	It("should detect <think> at end of prompt", func() {
+		Expect(DetectThinkingForcedOpen("Some prompt<think>")).To(BeTrue())
+		Expect(DetectThinkingForcedOpen("Some prompt<think>\n")).To(BeTrue())
+	})
+
+	It("should detect <thinking> at end of prompt", func() {
+		Expect(DetectThinkingForcedOpen("Some prompt<thinking>")).To(BeTrue())
+		Expect(DetectThinkingForcedOpen("Some prompt<thinking>\n")).To(BeTrue())
+	})
+
+	It("should detect model-specific tags", func() {
+		Expect(DetectThinkingForcedOpen("Some prompt<|inner_prefix|>")).To(BeTrue())
+		Expect(DetectThinkingForcedOpen("Some prompt<|START_THINKING|>")).To(BeTrue())
+		Expect(DetectThinkingForcedOpen("Some prompt<seed:think>")).To(BeTrue())
+		Expect(DetectThinkingForcedOpen("Some prompt[THINK]")).To(BeTrue())
+		Expect(DetectThinkingForcedOpen("Some prompt[THINK]\n")).To(BeTrue())
+	})
+
+	It("should not detect if tag is in the middle", func() {
+		Expect(DetectThinkingForcedOpen("Some <think> prompt")).To(BeFalse())
+		Expect(DetectThinkingForcedOpen("<think>reasoning</think>")).To(BeFalse())
+	})
+
+	It("should not detect if no thinking tag", func() {
+		Expect(DetectThinkingForcedOpen("Some regular prompt")).To(BeFalse())
+		Expect(DetectThinkingForcedOpen("")).To(BeFalse())
+	})
+})
+
+var _ = Describe("Extract", func() {
+	Context("when content has no reasoning tags", func() {
+		It("should return empty reasoning and original content", func() {
+			content := "This is regular content without any tags."
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal(content))
+		})
+
+		It("should handle empty string", func() {
+			content := ""
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(BeEmpty())
+		})
+
+		It("should handle content with only whitespace", func() {
+			content := "   \n\t  "
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal(content))
+		})
+	})
+
+	Context("when content has <thinking> tags", func() {
+		It("should extract reasoning from single thinking block", func() {
+			content := "Some text <thinking>This is my reasoning</thinking> More text"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("This is my reasoning"))
+			Expect(cleaned).To(Equal("Some text  More text"))
+		})
+
+		It("should extract reasoning and preserve surrounding content", func() {
+			content := "Before <thinking>Reasoning here</thinking> After"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Reasoning here"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should handle thinking block at the start", func() {
+			content := "<thinking>Start reasoning</thinking> Regular content"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Start reasoning"))
+			Expect(cleaned).To(Equal(" Regular content"))
+		})
+
+		It("should handle thinking block at the end", func() {
+			content := "Regular content <thinking>End reasoning</thinking>"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("End reasoning"))
+			Expect(cleaned).To(Equal("Regular content "))
+		})
+
+		It("should handle only thinking block", func() {
+			content := "<thinking>Only reasoning</thinking>"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Only reasoning"))
+			Expect(cleaned).To(BeEmpty())
+		})
+
+		It("should trim whitespace from reasoning content", func() {
+			content := "Text <thinking>  \n  Reasoning with spaces  \n  </thinking> More"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Reasoning with spaces"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+
+	Context("when content has <think> tags", func() {
+		It("should extract reasoning from redacted_reasoning block", func() {
+			content := "Text <think>Redacted reasoning</think> More"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Redacted reasoning"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+
+		It("should handle redacted_reasoning with multiline content", func() {
+			content := "Before <think>Line 1\nLine 2\nLine 3</think> After"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Line 1\nLine 2\nLine 3"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should handle redacted_reasoning with complex content", func() {
+			content := "Start <think>Complex reasoning\nwith\nmultiple\nlines</think> End"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Complex reasoning\nwith\nmultiple\nlines"))
+			Expect(cleaned).To(Equal("Start  End"))
+		})
+	})
+
+	Context("when content has multiple reasoning blocks", func() {
+		It("should concatenate multiple thinking blocks with newlines", func() {
+			content := "Text <thinking>First</thinking> Middle <thinking>Second</thinking> End"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("First\n\nSecond"))
+			Expect(cleaned).To(Equal("Text  Middle  End"))
+		})
+
+		It("should handle multiple different tag types", func() {
+			content := "A <thinking>One</thinking> B <think>Two</think> C <think>Three</think> D"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(ContainSubstring("One"))
+			Expect(reasoning).To(ContainSubstring("Two"))
+			Expect(reasoning).To(ContainSubstring("Three"))
+			Expect(cleaned).To(Equal("A  B  C  D"))
+		})
+
+		It("should handle nested tags correctly (extracts first match)", func() {
+			content := "Text <thinking>Outer <think>Inner</think></thinking> More"
+			reasoning, cleaned := Extract(content)
+			// Should extract the outer thinking block
+			Expect(reasoning).To(ContainSubstring("Outer"))
+			Expect(reasoning).To(ContainSubstring("Inner"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+
+	Context("when content has unclosed reasoning tags", func() {
+		It("should extract unclosed thinking block", func() {
+			content := "Text <thinking>Unclosed reasoning"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Unclosed reasoning"))
+			Expect(cleaned).To(Equal("Text "))
+		})
+
+		It("should extract unclosed think block", func() {
+			content := "Before <think>Incomplete"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Incomplete"))
+			Expect(cleaned).To(Equal("Before "))
+		})
+
+		It("should extract unclosed redacted_reasoning block", func() {
+			content := "Start <think>Partial reasoning content"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Partial reasoning content"))
+			Expect(cleaned).To(Equal("Start "))
+		})
+
+		It("should handle unclosed tag at the end", func() {
+			content := "Regular content <thinking>Unclosed at end"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Unclosed at end"))
+			Expect(cleaned).To(Equal("Regular content "))
+		})
+	})
+
+	Context("when content has empty reasoning blocks", func() {
+		It("should ignore empty thinking block", func() {
+			content := "Text <thinking></thinking> More"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+
+		It("should ignore thinking block with only whitespace", func() {
+			content := "Text <thinking>   \n\t  </thinking> More"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+
+	Context("when content has reasoning tags with special characters", func() {
+		It("should handle reasoning with newlines", func() {
+			content := "Before <thinking>Line 1\nLine 2\nLine 3</thinking> After"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Line 1\nLine 2\nLine 3"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should handle reasoning with code blocks", func() {
+			content := "Text <thinking>Reasoning with ```code``` blocks</thinking> More"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Reasoning with ```code``` blocks"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+
+		It("should handle reasoning with JSON", func() {
+			content := "Before <think>{\"key\": \"value\"}</think> After"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("{\"key\": \"value\"}"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should handle reasoning with HTML-like content", func() {
+			content := "Text <thinking>Reasoning with <tags> inside</thinking> More"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Reasoning with <tags> inside"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+
+	Context("when content has reasoning mixed with regular content", func() {
+		It("should preserve content order correctly", func() {
+			content := "Start <thinking>Reasoning</thinking> Middle <think>More reasoning</think> End"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(ContainSubstring("Reasoning"))
+			Expect(reasoning).To(ContainSubstring("More reasoning"))
+			Expect(cleaned).To(Equal("Start  Middle  End"))
+		})
+
+		It("should handle reasoning in the middle of a sentence", func() {
+			content := "This is a <thinking>reasoning</thinking> sentence."
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("reasoning"))
+			Expect(cleaned).To(Equal("This is a  sentence."))
+		})
+	})
+
+	Context("edge cases without WithThinkingForcedOpen", func() {
+		It("should handle content with only opening tag", func() {
+			content := "<thinking>"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal(""))
+		})
+
+		It("should handle content with only closing tag (no content before)", func() {
+			content := "</thinking>"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(BeEmpty())
+		})
+
+		It("should extract reasoning when only closing tag is present", func() {
+			// GLM-4 style: reasoning content followed by closing tag without opening tag
+			content := "This is reasoning content</think>this is the actual response"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("This is reasoning content"))
+			Expect(cleaned).To(Equal("this is the actual response"))
+		})
+
+		It("should handle closing-only tag with multiline reasoning", func() {
+			content := "1. First point\n2. Second point\n3. Third point</think>Final answer"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("1. First point\n2. Second point\n3. Third point"))
+			Expect(cleaned).To(Equal("Final answer"))
+		})
+
+		It("should handle closing-only tag with complex reasoning (GLM-4 example)", func() {
+			content := "**Analyze the user's input:** The user says something.\n\n**Final Decision:** Output the text.</think>this is a test"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("**Analyze the user's input:** The user says something.\n\n**Final Decision:** Output the text."))
+			Expect(cleaned).To(Equal("this is a test"))
+		})
+
+		It("should handle closing-only thinking tag", func() {
+			content := "Some reasoning here</thinking>actual content"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Some reasoning here"))
+			Expect(cleaned).To(Equal("actual content"))
+		})
+
+		It("should handle mismatched tags", func() {
+			content := "<thinking>Content</think>"
+			reasoning, cleaned := Extract(content)
+			// Should extract unclosed thinking block
+			Expect(reasoning).To(ContainSubstring("Content"))
+			Expect(cleaned).To(Equal(""))
+		})
+
+		It("should handle very long reasoning content", func() {
+			longReasoning := strings.Repeat("This is reasoning content. ", 100)
+			content := "Text <thinking>" + longReasoning + "</thinking> More"
+			reasoning, cleaned := Extract(content)
+			// TrimSpace is applied, so we need to account for that
+			Expect(reasoning).To(Equal(strings.TrimSpace(longReasoning)))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+
+		It("should handle reasoning with unicode characters", func() {
+			content := "Text <thinking>Reasoning with 中文 and emoji 🧠</thinking> More"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Reasoning with 中文 and emoji 🧠"))
+			Expect(cleaned).To(Equal("Text  More"))
+		})
+	})
+
+	Context("with WithThinkingForcedOpen option", func() {
+		It("should treat all content as reasoning until closing tag", func() {
+			content := "This is reasoning</think>this is content"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("This is reasoning"))
+			Expect(cleaned).To(Equal("this is content"))
+		})
+
+		It("should treat all content as reasoning when no closing tag (streaming)", func() {
+			content := "This is reasoning content still streaming"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("This is reasoning content still streaming"))
+			Expect(cleaned).To(BeEmpty())
+		})
+
+		It("should handle GLM-4 style output", func() {
+			content := "**Analyze:** The user says something.\n\n**Final Decision:** Output the text.</think>this is a test"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("**Analyze:** The user says something.\n\n**Final Decision:** Output the text."))
+			Expect(cleaned).To(Equal("this is a test"))
+		})
+
+		It("should handle multiline reasoning with closing tag", func() {
+			content := "1. First point\n2. Second point\n3. Third point</think>Final answer"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("1. First point\n2. Second point\n3. Third point"))
+			Expect(cleaned).To(Equal("Final answer"))
+		})
+
+		It("should handle </thinking> closing tag", func() {
+			content := "Some reasoning here</thinking>actual content"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("Some reasoning here"))
+			Expect(cleaned).To(Equal("actual content"))
+		})
+
+		It("should handle additional reasoning blocks after initial forced open", func() {
+			content := "Initial reasoning</think>content<think>more reasoning</think>final content"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("Initial reasoning\n\nmore reasoning"))
+			Expect(cleaned).To(Equal("contentfinal content"))
+		})
+
+		It("should handle empty content", func() {
+			reasoning, cleaned := Extract("", WithThinkingForcedOpen())
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(BeEmpty())
+		})
+
+		It("should handle only closing tag", func() {
+			content := "</think>only content"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(BeEmpty())
+			Expect(cleaned).To(Equal("only content"))
+		})
+
+		It("should find earliest closing tag", func() {
+			// </think> comes before </thinking>
+			content := "Reasoning</think>content</thinking>more"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("Reasoning"))
+			Expect(cleaned).To(Equal("content</thinking>more"))
+		})
+
+		It("should handle Command R7B closing tag", func() {
+			content := "Reasoning content<|END_THINKING|>actual response"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("Reasoning content"))
+			Expect(cleaned).To(Equal("actual response"))
+		})
+
+		It("should handle Apertus closing tag", func() {
+			content := "Reasoning content<|inner_suffix|>actual response"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("Reasoning content"))
+			Expect(cleaned).To(Equal("actual response"))
+		})
+
+		It("should handle Seed closing tag", func() {
+			content := "Reasoning content</seed:think>actual response"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("Reasoning content"))
+			Expect(cleaned).To(Equal("actual response"))
+		})
+
+		It("should handle Magistral closing tag", func() {
+			content := "Reasoning content[/THINK]actual response"
+			reasoning, cleaned := Extract(content, WithThinkingForcedOpen())
+			Expect(reasoning).To(Equal("Reasoning content"))
+			Expect(cleaned).To(Equal("actual response"))
+		})
+	})
+
+	Context("with model-specific tag pairs", func() {
+		It("should extract Command R7B reasoning tags", func() {
+			content := "Before <|START_THINKING|>reasoning here<|END_THINKING|> After"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("reasoning here"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should extract Apertus reasoning tags", func() {
+			content := "Before <|inner_prefix|>reasoning here<|inner_suffix|> After"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("reasoning here"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should extract Seed reasoning tags", func() {
+			content := "Before <seed:think>reasoning here</seed:think> After"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("reasoning here"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should extract Magistral reasoning tags", func() {
+			content := "Before [THINK]reasoning here[/THINK] After"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("reasoning here"))
+			Expect(cleaned).To(Equal("Before  After"))
+		})
+
+		It("should handle unclosed Command R7B tag", func() {
+			content := "Before <|START_THINKING|>reasoning still streaming"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("reasoning still streaming"))
+			Expect(cleaned).To(Equal("Before "))
+		})
+
+		It("should handle unclosed Apertus tag", func() {
+			content := "Before <|inner_prefix|>reasoning still streaming"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("reasoning still streaming"))
+			Expect(cleaned).To(Equal("Before "))
+		})
+
+		It("should handle unclosed Seed tag", func() {
+			content := "Before <seed:think>reasoning still streaming"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("reasoning still streaming"))
+			Expect(cleaned).To(Equal("Before "))
+		})
+
+		It("should handle unclosed Magistral tag", func() {
+			content := "Before [THINK]reasoning still streaming"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("reasoning still streaming"))
+			Expect(cleaned).To(Equal("Before "))
+		})
+
+		It("should handle closing-only Command R7B tag", func() {
+			content := "Reasoning content<|END_THINKING|>actual response"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Reasoning content"))
+			Expect(cleaned).To(Equal("actual response"))
+		})
+
+		It("should handle closing-only Apertus tag", func() {
+			content := "Reasoning content<|inner_suffix|>actual response"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Reasoning content"))
+			Expect(cleaned).To(Equal("actual response"))
+		})
+
+		It("should handle closing-only Seed tag", func() {
+			content := "Reasoning content</seed:think>actual response"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Reasoning content"))
+			Expect(cleaned).To(Equal("actual response"))
+		})
+
+		It("should handle closing-only Magistral tag", func() {
+			content := "Reasoning content[/THINK]actual response"
+			reasoning, cleaned := Extract(content)
+			Expect(reasoning).To(Equal("Reasoning content"))
+			Expect(cleaned).To(Equal("actual response"))
+		})
+	})
+})