|
| 1 | +package ensembleserver |
| 2 | + |
| 3 | +import ( |
| 4 | + "encoding/json" |
| 5 | + "fmt" |
| 6 | + "io" |
| 7 | + "net/http" |
| 8 | + "strconv" |
| 9 | + "strings" |
| 10 | + "time" |
| 11 | + |
| 12 | + "github.com/vllm-project/semantic-router/src/semantic-router/pkg/config" |
| 13 | + "github.com/vllm-project/semantic-router/src/semantic-router/pkg/ensemble" |
| 14 | + "github.com/vllm-project/semantic-router/src/semantic-router/pkg/headers" |
| 15 | + "github.com/vllm-project/semantic-router/src/semantic-router/pkg/observability/logging" |
| 16 | +) |
| 17 | + |
| 18 | +// EnsembleServer handles OpenAI-compatible ensemble requests |
| 19 | +type EnsembleServer struct { |
| 20 | + factory *ensemble.Factory |
| 21 | + config *config.RouterConfig |
| 22 | +} |
| 23 | + |
| 24 | +// Init starts the ensemble API server |
| 25 | +func Init(cfg *config.RouterConfig, port int) error { |
| 26 | + if cfg == nil { |
| 27 | + return fmt.Errorf("configuration not initialized") |
| 28 | + } |
| 29 | + |
| 30 | + if !cfg.Ensemble.Enabled { |
| 31 | + logging.Infof("Ensemble service is disabled in configuration") |
| 32 | + return nil |
| 33 | + } |
| 34 | + |
| 35 | + // Initialize ensemble factory |
| 36 | + ensembleConfig := &ensemble.Config{ |
| 37 | + Enabled: cfg.Ensemble.Enabled, |
| 38 | + DefaultStrategy: ensemble.Strategy(cfg.Ensemble.DefaultStrategy), |
| 39 | + DefaultMinResponses: cfg.Ensemble.DefaultMinResponses, |
| 40 | + TimeoutSeconds: cfg.Ensemble.TimeoutSeconds, |
| 41 | + MaxConcurrentRequests: cfg.Ensemble.MaxConcurrentRequests, |
| 42 | + } |
| 43 | + factory := ensemble.NewFactory(ensembleConfig) |
| 44 | + |
| 45 | + // Register endpoint mappings from config |
| 46 | + for modelName, endpoint := range cfg.Ensemble.EndpointMappings { |
| 47 | + factory.RegisterEndpoint(modelName, endpoint) |
| 48 | + } |
| 49 | + |
| 50 | + server := &EnsembleServer{ |
| 51 | + factory: factory, |
| 52 | + config: cfg, |
| 53 | + } |
| 54 | + |
| 55 | + // Create HTTP server |
| 56 | + mux := server.setupRoutes() |
| 57 | + httpServer := &http.Server{ |
| 58 | + Addr: fmt.Sprintf(":%d", port), |
| 59 | + Handler: mux, |
| 60 | + ReadTimeout: 60 * time.Second, |
| 61 | + WriteTimeout: 60 * time.Second, |
| 62 | + IdleTimeout: 120 * time.Second, |
| 63 | + } |
| 64 | + |
| 65 | + logging.Infof("Ensemble API server listening on port %d", port) |
| 66 | + return httpServer.ListenAndServe() |
| 67 | +} |
| 68 | + |
| 69 | +// setupRoutes configures HTTP routes |
| 70 | +func (s *EnsembleServer) setupRoutes() *http.ServeMux { |
| 71 | + mux := http.NewServeMux() |
| 72 | + |
| 73 | + // OpenAI-compatible endpoints |
| 74 | + mux.HandleFunc("/v1/chat/completions", s.handleChatCompletions) |
| 75 | + mux.HandleFunc("/health", s.handleHealth) |
| 76 | + |
| 77 | + return mux |
| 78 | +} |
| 79 | + |
| 80 | +// handleHealth returns service health status |
| 81 | +func (s *EnsembleServer) handleHealth(w http.ResponseWriter, r *http.Request) { |
| 82 | + w.Header().Set("Content-Type", "application/json") |
| 83 | + w.WriteHeader(http.StatusOK) |
| 84 | + json.NewEncoder(w).Encode(map[string]interface{}{ |
| 85 | + "status": "healthy", |
| 86 | + "service": "ensemble", |
| 87 | + }) |
| 88 | +} |
| 89 | + |
| 90 | +// handleChatCompletions processes OpenAI chat completion requests with ensemble |
| 91 | +func (s *EnsembleServer) handleChatCompletions(w http.ResponseWriter, r *http.Request) { |
| 92 | + if r.Method != http.MethodPost { |
| 93 | + http.Error(w, "Method not allowed", http.StatusMethodNotAllowed) |
| 94 | + return |
| 95 | + } |
| 96 | + |
| 97 | + // Read request body |
| 98 | + body, err := io.ReadAll(r.Body) |
| 99 | + if err != nil { |
| 100 | + logging.Errorf("Failed to read request body: %v", err) |
| 101 | + http.Error(w, "Failed to read request body", http.StatusBadRequest) |
| 102 | + return |
| 103 | + } |
| 104 | + defer r.Body.Close() |
| 105 | + |
| 106 | + // Parse ensemble headers |
| 107 | + ensembleEnabled := strings.ToLower(r.Header.Get(headers.EnsembleEnable)) == "true" |
| 108 | + if !ensembleEnabled { |
| 109 | + http.Error(w, "Ensemble not enabled in request headers", http.StatusBadRequest) |
| 110 | + return |
| 111 | + } |
| 112 | + |
| 113 | + // Parse models list |
| 114 | + modelsHeader := r.Header.Get(headers.EnsembleModels) |
| 115 | + if modelsHeader == "" { |
| 116 | + http.Error(w, "No models specified in ensemble header", http.StatusBadRequest) |
| 117 | + return |
| 118 | + } |
| 119 | + |
| 120 | + var models []string |
| 121 | + for _, model := range strings.Split(modelsHeader, ",") { |
| 122 | + trimmedModel := strings.TrimSpace(model) |
| 123 | + if trimmedModel != "" { |
| 124 | + models = append(models, trimmedModel) |
| 125 | + } |
| 126 | + } |
| 127 | + |
| 128 | + if len(models) == 0 { |
| 129 | + http.Error(w, "No valid models specified", http.StatusBadRequest) |
| 130 | + return |
| 131 | + } |
| 132 | + |
| 133 | + // Parse strategy |
| 134 | + strategy := ensemble.Strategy(r.Header.Get(headers.EnsembleStrategy)) |
| 135 | + if strategy == "" { |
| 136 | + strategy = s.factory.GetDefaultStrategy() |
| 137 | + } |
| 138 | + |
| 139 | + // Parse min responses |
| 140 | + minResponses := s.factory.GetDefaultMinResponses() |
| 141 | + if minRespHeader := r.Header.Get(headers.EnsembleMinResponses); minRespHeader != "" { |
| 142 | + if parsed, err := strconv.Atoi(minRespHeader); err == nil && parsed > 0 { |
| 143 | + minResponses = parsed |
| 144 | + } |
| 145 | + } |
| 146 | + |
| 147 | + logging.Infof("Ensemble request: models=%v, strategy=%s, minResponses=%d", models, strategy, minResponses) |
| 148 | + |
| 149 | + // Forward headers for authentication |
| 150 | + headerMap := make(map[string]string) |
| 151 | + for key, values := range r.Header { |
| 152 | + if len(values) > 0 { |
| 153 | + headerMap[key] = values[0] |
| 154 | + } |
| 155 | + } |
| 156 | + |
| 157 | + // Build ensemble request |
| 158 | + ensembleReq := &ensemble.Request{ |
| 159 | + Models: models, |
| 160 | + Strategy: strategy, |
| 161 | + MinResponses: minResponses, |
| 162 | + OriginalRequest: body, |
| 163 | + Headers: headerMap, |
| 164 | + Context: r.Context(), |
| 165 | + } |
| 166 | + |
| 167 | + // Execute ensemble orchestration |
| 168 | + ensembleResp := s.factory.Execute(ensembleReq) |
| 169 | + |
| 170 | + // Check for errors |
| 171 | + if ensembleResp.Error != nil { |
| 172 | + logging.Errorf("Ensemble execution failed: %v", ensembleResp.Error) |
| 173 | + http.Error(w, fmt.Sprintf("Ensemble orchestration failed: %v", ensembleResp.Error), http.StatusInternalServerError) |
| 174 | + return |
| 175 | + } |
| 176 | + |
| 177 | + // Add ensemble metadata headers |
| 178 | + w.Header().Set(headers.VSREnsembleUsed, "true") |
| 179 | + w.Header().Set(headers.VSREnsembleModelsQueried, strconv.Itoa(ensembleResp.ModelsQueried)) |
| 180 | + w.Header().Set(headers.VSREnsembleResponsesReceived, strconv.Itoa(ensembleResp.ResponsesReceived)) |
| 181 | + w.Header().Set("Content-Type", "application/json") |
| 182 | + |
| 183 | + // Return the aggregated response |
| 184 | + logging.Infof("Ensemble execution successful: queried=%d, received=%d, strategy=%s", |
| 185 | + ensembleResp.ModelsQueried, ensembleResp.ResponsesReceived, ensembleResp.Strategy) |
| 186 | + |
| 187 | + w.WriteHeader(http.StatusOK) |
| 188 | + w.Write(ensembleResp.FinalResponse) |
| 189 | +} |
0 commit comments