feat: implement decision-based routing with plugin architecture

Signed-off-by: bitliu <[email protected]>

Author: Xunzhuo

deploy/kubernetes/ai-gateway/semantic-router-values/values.yaml

@@ -145,7 +145,8 @@ func (r *OpenAIRouter) handleAutoModelRouting(openAIRequest *openai.ChatCompleti
 	r.recordRoutingDecision(ctx, decisionName, originalModel, matchedModel, reasoningDecision)
 
 	// Track VSR decision information
-	r.trackVSRDecision(ctx, decisionName, matchedModel, reasoningDecision.UseReasoning)
+	// categoryName is already set in ctx.VSRSelectedCategory by performDecisionEvaluationAndModelSelection
+	r.trackVSRDecision(ctx, ctx.VSRSelectedCategory, decisionName, matchedModel, reasoningDecision.UseReasoning)
 
 	// Track model routing metrics
 	metrics.RecordModelRouting(originalModel, matchedModel)

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

@@ -34,12 +34,13 @@ type RequestContext struct {
 	TTFTSeconds  float64
 
 	// VSR decision tracking
-	VSRSelectedCategory     string           // The category selected by VSR
+	VSRSelectedCategory     string           // The category from domain classification (MMLU category)
+	VSRSelectedDecisionName string           // The decision name from DecisionEngine evaluation
 	VSRReasoningMode        string           // "on" or "off" - whether reasoning mode was determined to be used
 	VSRSelectedModel        string           // The model selected by VSR
 	VSRCacheHit             bool             // Whether this request hit the cache
 	VSRInjectedSystemPrompt bool             // Whether a system prompt was injected into the request
-	VSRSelectedDecision     *config.Decision // The decision selected by DecisionEngine
+	VSRSelectedDecision     *config.Decision // The decision object selected by DecisionEngine (for plugins)
 
 	// Tracing context
 	TraceContext context.Context // OpenTelemetry trace context for span propagation

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

@@ -54,7 +54,7 @@ func (r *OpenAIRouter) handleResponseHeaders(v *ext_proc.ProcessingRequest_Respo
 	if isSuccessful && !ctx.VSRCacheHit && ctx != nil {
 		var setHeaders []*core.HeaderValueOption
 
-		// Add x-vsr-selected-category header
+		// Add x-vsr-selected-category header (from domain classification)
 		if ctx.VSRSelectedCategory != "" {
 			setHeaders = append(setHeaders, &core.HeaderValueOption{
 				Header: &core.HeaderValue{
@@ -64,6 +64,16 @@ func (r *OpenAIRouter) handleResponseHeaders(v *ext_proc.ProcessingRequest_Respo
 			})
 		}
 
+		// Add x-vsr-selected-decision header (from decision evaluation)
+		if ctx.VSRSelectedDecisionName != "" {
+			setHeaders = append(setHeaders, &core.HeaderValueOption{
+				Header: &core.HeaderValue{
+					Key:      headers.VSRSelectedDecision,
+					RawValue: []byte(ctx.VSRSelectedDecisionName),
+				},
+			})
+		}
+
 		// Add x-vsr-selected-reasoning header
 		if ctx.VSRReasoningMode != "" {
 			setHeaders = append(setHeaders, &core.HeaderValueOption{

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

@@ -57,8 +57,11 @@ func (r *OpenAIRouter) recordRoutingDecision(ctx *RequestContext, decisionName s
 }
 
 // trackVSRDecision tracks VSR decision information in context
-func (r *OpenAIRouter) trackVSRDecision(ctx *RequestContext, decisionName string, matchedModel string, useReasoning bool) {
-	ctx.VSRSelectedCategory = decisionName
+// categoryName: the category from domain classification (MMLU category)
+// decisionName: the decision name from DecisionEngine evaluation
+func (r *OpenAIRouter) trackVSRDecision(ctx *RequestContext, categoryName string, decisionName string, matchedModel string, useReasoning bool) {
+	ctx.VSRSelectedCategory = categoryName
+	ctx.VSRSelectedDecisionName = decisionName
 	ctx.VSRSelectedModel = matchedModel
 	if useReasoning {
 		ctx.VSRReasoningMode = "on"

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

@@ -56,10 +56,23 @@ func (r *OpenAIRouter) performDecisionEvaluationAndModelSelection(originalModel
 	// Store the selected decision in context for later use (e.g., header mutations)
 	ctx.VSRSelectedDecision = result.Decision
 
+	// Extract domain category from matched rules (for VSRSelectedCategory header)
+	// MatchedRules contains rule names like "domain:math", "keyword:thinking", etc.
+	// We extract the first domain rule as the category
+	categoryName := ""
+	for _, rule := range result.MatchedRules {
+		if strings.HasPrefix(rule, "domain:") {
+			categoryName = strings.TrimPrefix(rule, "domain:")
+			break
+		}
+	}
+	// Store category in context for response headers
+	ctx.VSRSelectedCategory = categoryName
+
 	decisionName = result.Decision.Name
 	evaluationConfidence = result.Confidence
-	logging.Infof("Decision Evaluation Result: decision=%s, confidence=%.3f, matched_rules=%v",
-		decisionName, evaluationConfidence, result.MatchedRules)
+	logging.Infof("Decision Evaluation Result: decision=%s, category=%s, confidence=%.3f, matched_rules=%v",
+		decisionName, categoryName, evaluationConfidence, result.MatchedRules)
 
 	// Select best model from the decision's ModelRefs
 	if len(result.Decision.ModelRefs) > 0 {

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

@@ -24,10 +24,16 @@ const (
 // Vector Semantic Router decision-making information for debugging and monitoring.
 // Headers are only added when the request is successful and did not hit the cache.
 const (
-	// VSRSelectedCategory indicates the category selected by VSR during classification.
+	// VSRSelectedCategory indicates the category selected by VSR during domain classification.
+	// This comes from the domain classifier (MMLU categories).
 	// Example values: "math", "business", "biology", "computer_science"
 	VSRSelectedCategory = "x-vsr-selected-category"
 
+	// VSRSelectedDecision indicates the decision selected by VSR during decision evaluation.
+	// This is the final routing decision made by the DecisionEngine.
+	// Example values: "math_decision", "business_decision", "thinking_decision"
+	VSRSelectedDecision = "x-vsr-selected-decision"
+
 	// VSRSelectedReasoning indicates whether reasoning mode was determined to be used.
 	// Values: "on" (reasoning enabled) or "off" (reasoning disabled)
 	VSRSelectedReasoning = "x-vsr-selected-reasoning"