fix: don't set reasoning effort for non-reasoning models

Signed-off-by: Huamin Chen <[email protected]>

Author: rootfs

config/config.yaml

@@ -253,4 +253,32 @@ api:
       sample_rate: 1.0                   # Collect metrics for all requests (1.0 = 100%, 0.5 = 50%)
       # Histogram buckets for metrics (directly configure what you need)
       duration_buckets: [0.001, 0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1, 2.5, 5, 10, 30]
-      size_buckets: [1, 2, 5, 10, 20, 50, 100, 200] 
+      size_buckets: [1, 2, 5, 10, 20, 50, 100, 200]
+
+# Model reasoning configurations - define how different models handle reasoning syntax
+model_reasoning_configs:
+  - name: "deepseek"
+    patterns: ["deepseek", "ds-", "ds_", "ds:", "ds "]
+    reasoning_syntax:
+      type: "chat_template_kwargs"
+      parameter: "thinking"
+  
+  - name: "qwen3"
+    patterns: ["qwen3"]
+    reasoning_syntax:
+      type: "chat_template_kwargs"
+      parameter: "enable_thinking"
+  
+  - name: "gpt-oss"
+    patterns: ["gpt-oss", "gpt_oss"]
+    reasoning_syntax:
+      type: "reasoning_effort"
+      parameter: "reasoning_effort"
+  
+  - name: "gpt"
+    patterns: ["gpt"]
+    reasoning_syntax:
+      type: "reasoning_effort"
+      parameter: "reasoning_effort"
+  
+# No default fallback - unknown models should have no reasoning syntax applied 

src/semantic-router/pkg/config/config.go

@@ -4,6 +4,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"strings"
 	"sync"
 
 	"gopkg.in/yaml.v3"
@@ -45,6 +46,9 @@ type RouterConfig struct {
 	// Default reasoning effort level (low, medium, high) when not specified per category
 	DefaultReasoningEffort string `yaml:"default_reasoning_effort,omitempty"`
 
+	// Model reasoning configurations to define how different models handle reasoning syntax
+	ModelReasoningConfigs []ModelReasoningConfig `yaml:"model_reasoning_configs,omitempty"`
+
 	// Semantic cache configuration
 	SemanticCache SemanticCacheConfig `yaml:"semantic_cache"`
 
@@ -264,6 +268,73 @@ type Category struct {
 	ModelScores          []ModelScore `yaml:"model_scores"`
 }
 
+// ModelReasoningSyntax defines how a model handles reasoning mode
+type ModelReasoningSyntax struct {
+	Type      string `yaml:"type"`      // "chat_template_kwargs" or "reasoning_effort"
+	Parameter string `yaml:"parameter"` // "thinking", "enable_thinking", "reasoning_effort", etc.
+}
+
+// ModelReasoningConfig defines reasoning configuration for a model family
+type ModelReasoningConfig struct {
+	Name            string               `yaml:"name"`
+	Patterns        []string             `yaml:"patterns"` // patterns to match against model names
+	ReasoningSyntax ModelReasoningSyntax `yaml:"reasoning_syntax"`
+}
+
+// FindModelReasoningConfig finds the appropriate reasoning configuration for a given model name
+func (rc *RouterConfig) FindModelReasoningConfig(modelName string) *ModelReasoningConfig {
+	if rc == nil || len(rc.ModelReasoningConfigs) == 0 {
+		return nil
+	}
+
+	modelLower := strings.ToLower(strings.TrimSpace(modelName))
+
+	// Look for exact pattern matches first
+	for _, config := range rc.ModelReasoningConfigs {
+		for _, pattern := range config.Patterns {
+			if matchesPattern(modelLower, strings.ToLower(pattern)) {
+				return &config
+			}
+		}
+	}
+
+	// Look for a default/fallback configuration
+	for _, config := range rc.ModelReasoningConfigs {
+		if config.Name == "default" || contains(config.Patterns, "*") {
+			return &config
+		}
+	}
+
+	return nil
+}
+
+// matchesPattern checks if a model name matches a given pattern
+func matchesPattern(modelName, pattern string) bool {
+	if pattern == "*" {
+		return true
+	}
+
+	// Handle prefix patterns like "ds-", "ds_", "ds:", "ds "
+	if strings.HasSuffix(pattern, "-") || strings.HasSuffix(pattern, "_") ||
+		strings.HasSuffix(pattern, ":") || strings.HasSuffix(pattern, " ") {
+		return strings.HasPrefix(modelName, pattern) ||
+			(len(modelName) >= len(pattern)-1 && modelName == pattern[:len(pattern)-1])
+	}
+
+	// Handle exact matches and contains
+	return strings.Contains(modelName, pattern)
+}
+
+// contains checks if a slice contains a string
+func contains(slice []string, item string) bool {
+	for _, s := range slice {
+		if s == item {
+			return true
+		}
+	}
+	return false
+}
+
 var (
 	config     *RouterConfig
 	configOnce sync.Once

src/semantic-router/pkg/extproc/reason_mode_selector.go

@@ -6,6 +6,7 @@ import (
 	"log"
 	"strings"
 
+	"github.com/vllm-project/semantic-router/semantic-router/pkg/config"
 	"github.com/vllm-project/semantic-router/semantic-router/pkg/metrics"
 )
 
@@ -47,61 +48,42 @@ func (r *OpenAIRouter) getReasoningModeAndCategory(query string) (bool, string)
 	return false, categoryName
 }
 
-// hasDeepSeekAlias returns true if the model uses a short alias for DeepSeek (e.g., "ds-*")
-// Rules:
-//   - Accept only when the model string starts with: "ds-", "ds_", "ds:", "ds " or exactly equals "ds"
-//   - Do NOT match occurrences of "ds" in the middle of the model name (e.g., "foo-ds-1b")
-func hasDeepSeekAlias(lower string) bool {
-	lower = strings.TrimSpace(lower)
-	if strings.HasPrefix(lower, "ds") {
-		if len(lower) == 2 { // exactly "ds"
-			return true
-		}
-		sep := lower[2]
-		return sep == '-' || sep == '_' || sep == ':' || sep == ' '
+// getModelReasoningConfig finds the reasoning configuration for a model using the config system
+func (r *OpenAIRouter) getModelReasoningConfig(model string) *config.ModelReasoningConfig {
+	if r.Config == nil {
+		return nil
 	}
-	return false
+	return r.Config.FindModelReasoningConfig(model)
 }
 
-// getModelFamilyAndTemplateParam returns a normalized model family name and the template param to be used (if any)
-func getModelFamilyAndTemplateParam(model string) (string, string) {
-	lower := strings.ToLower(strings.TrimSpace(model))
-	if strings.Contains(lower, "qwen3") {
-		return "qwen3", "enable_thinking"
-	}
-	if strings.Contains(lower, "deepseek") || hasDeepSeekAlias(lower) {
-		return "deepseek", "thinking"
+// buildReasoningRequestFields returns the appropriate fields to add to the request based on model config
+func (r *OpenAIRouter) buildReasoningRequestFields(model string, useReasoning bool, categoryName string) (map[string]interface{}, string) {
+	modelConfig := r.getModelReasoningConfig(model)
+	if modelConfig == nil {
+		// No configuration found for this model - don't apply any reasoning syntax
+		// Unknown models should not have reasoning fields added
+		return nil, "N/A"
 	}
-	// GPT-OSS family and generic GPT fall back to using reasoning_effort (OpenAI-compatible field)
-	if strings.Contains(lower, "gpt-oss") || strings.Contains(lower, "gpt_oss") {
-		return "gpt-oss", "reasoning_effort"
-	}
-	if strings.Contains(lower, "gpt") {
-		return "gpt", "reasoning_effort"
-	}
-	return "unknown", ""
-}
 
-// getChatTemplateKwargs returns the appropriate chat template kwargs based on model and reasoning mode
-func getChatTemplateKwargs(model string, useReasoning bool) map[string]interface{} {
-	lower := strings.ToLower(strings.TrimSpace(model))
-
-	// Qwen3: use enable_thinking true/false
-	if strings.Contains(lower, "qwen3") {
-		return map[string]interface{}{
-			"enable_thinking": useReasoning,
-		}
+	if !useReasoning {
+		// When reasoning is disabled, don't add any reasoning fields
+		return nil, "N/A"
 	}
 
-	// DeepSeek v3 family: use thinking true/false
-	if strings.Contains(lower, "deepseek") || strings.Contains(lower, "ds") {
-		return map[string]interface{}{
-			"thinking": useReasoning,
+	// When reasoning is enabled, use the configured syntax
+	switch modelConfig.ReasoningSyntax.Type {
+	case "chat_template_kwargs":
+		kwargs := map[string]interface{}{
+			modelConfig.ReasoningSyntax.Parameter: useReasoning,
 		}
+		return map[string]interface{}{"chat_template_kwargs": kwargs}, "N/A"
+	case "reasoning_effort":
+		effort := r.getReasoningEffort(categoryName)
+		return map[string]interface{}{"reasoning_effort": effort}, effort
+	default:
+		// Unknown reasoning syntax type - don't apply anything
+		return nil, "N/A"
 	}
-
-	// Default: no chat template kwargs for unknown models
-	return nil
 }
 
 // setReasoningModeToRequestBody adds chat_template_kwargs to the JSON request body
@@ -120,49 +102,60 @@ func (r *OpenAIRouter) setReasoningModeToRequestBody(requestBody []byte, enabled
 		}
 	}
 
-	family, param := getModelFamilyAndTemplateParam(model)
-
-	// Add chat_template_kwargs for reasoning mode
-	kwargs := getChatTemplateKwargs(model, enabled)
-	if kwargs != nil {
-		requestMap["chat_template_kwargs"] = kwargs
-	} else {
-		delete(requestMap, "chat_template_kwargs")
-	}
-	// Also set Reasoning-Effort in openai request
-	// This is a hack to get the reasoning mode for openai/gpt-oss-20b to work
-	originalReasoningEffort, ok := requestMap["reasoning_effort"]
-	if !ok {
-		// This seems to be the default for openai/gpt-oss models
-		originalReasoningEffort = "low"
+	// Get original reasoning effort for potential preservation
+	originalReasoningEffort, hasOriginalEffort := requestMap["reasoning_effort"]
+	if !hasOriginalEffort {
+		originalReasoningEffort = "low" // Default for compatibility
 	}
-	var appliedEffort string
+
+	// Clear existing reasoning-related fields
+	delete(requestMap, "chat_template_kwargs")
+	delete(requestMap, "reasoning_effort")
+
+	var appliedEffort string = "N/A"
+
 	if enabled {
-		// Use configurable reasoning effort based on category
-		effort := r.getReasoningEffort(categoryName)
-		requestMap["reasoning_effort"] = effort
+		// When reasoning is enabled, build the appropriate fields
+		reasoningFields, effort := r.buildReasoningRequestFields(model, enabled, categoryName)
+		for key, value := range reasoningFields {
+			requestMap[key] = value
+		}
 		appliedEffort = effort
 	} else {
-		requestMap["reasoning_effort"] = originalReasoningEffort
-		if s, ok := originalReasoningEffort.(string); ok {
-			appliedEffort = s
+		// When reasoning is disabled, only preserve reasoning_effort for gpt-oss models
+		modelConfig := r.getModelReasoningConfig(model)
+		if modelConfig != nil && modelConfig.Name == "gpt-oss" {
+			requestMap["reasoning_effort"] = originalReasoningEffort
+			if s, ok := originalReasoningEffort.(string); ok {
+				appliedEffort = s
+			}
 		}
+		// For all other models, reasoning fields remain cleared
 	}
 
 	log.Printf("Original reasoning effort: %s", originalReasoningEffort)
-	log.Printf("Added reasoning mode (enabled: %v) and reasoning effort (%s) to request for model: %s", enabled, requestMap["reasoning_effort"], model)
+	log.Printf("Added reasoning mode (enabled: %v) and reasoning effort (%s) to request for model: %s", enabled, appliedEffort, model)
 
 	// Record metrics for template usage and effort when enabled
 	if enabled {
-		// If we applied a known template param, record its usage
-		if kwargs != nil && param != "" {
-			metrics.RecordReasoningTemplateUsage(family, param)
-		} else if kwargs == nil && param == "reasoning_effort" {
-			// For GPT/GPT-OSS, we only set reasoning_effort
-			metrics.RecordReasoningTemplateUsage(family, param)
+		modelConfig := r.getModelReasoningConfig(model)
+		modelFamily := "unknown"
+		templateParam := "reasoning_effort" // default fallback
+
+		if modelConfig != nil {
+			modelFamily = modelConfig.Name
+			if modelConfig.ReasoningSyntax.Type == "chat_template_kwargs" {
+				templateParam = modelConfig.ReasoningSyntax.Parameter
+			} else {
+				templateParam = "reasoning_effort"
+			}
+		}
+
+		// Record template usage and effort
+		metrics.RecordReasoningTemplateUsage(modelFamily, templateParam)
+		if appliedEffort != "N/A" {
+			metrics.RecordReasoningEffortUsage(modelFamily, appliedEffort)
 		}
-		// Record which effort level was used for this family
-		metrics.RecordReasoningEffortUsage(family, appliedEffort)
 	}
 
 	// Serialize back to JSON