feat: 🎉 Vectors na Vectors, the begining

Translate 1536 <-> 768 , 3072 <-> 2048

pkg/metrics/metrics.go

@@ -0,0 +1,112 @@
+package metrics
+
+import (
+	"github.com/prometheus/client_golang/prometheus"
+)
+
+// Registry holds all vecna Prometheus metrics on a dedicated (non-global) registry.
+type Registry struct {
+	reg *prometheus.Registry
+
+	RequestsTotal       *prometheus.CounterVec
+	RequestDuration     *prometheus.HistogramVec
+	ForwardDuration     *prometheus.HistogramVec
+	TranslateDuration   *prometheus.HistogramVec
+	EndpointPriority    *prometheus.GaugeVec
+	EndpointInflight    *prometheus.GaugeVec
+	EndpointErrorsTotal *prometheus.CounterVec
+	TokensTotal         *prometheus.CounterVec
+}
+
+// New creates and registers all metrics on a fresh Prometheus registry.
+func New() *Registry {
+	reg := prometheus.NewRegistry()
+
+	r := &Registry{
+		reg: reg,
+
+		RequestsTotal: prometheus.NewCounterVec(prometheus.CounterOpts{
+			Name: "vecna_requests_total",
+			Help: "Total number of requests served by vecna.",
+		}, []string{"endpoint", "status"}),
+
+		RequestDuration: prometheus.NewHistogramVec(prometheus.HistogramOpts{
+			Name:    "vecna_request_duration_seconds",
+			Help:    "Total request wall-clock time.",
+			Buckets: prometheus.DefBuckets,
+		}, []string{"endpoint"}),
+
+		ForwardDuration: prometheus.NewHistogramVec(prometheus.HistogramOpts{
+			Name:    "vecna_forward_duration_seconds",
+			Help:    "Time spent waiting on the backing embedding model.",
+			Buckets: prometheus.DefBuckets,
+		}, []string{"target", "url"}),
+
+		TranslateDuration: prometheus.NewHistogramVec(prometheus.HistogramOpts{
+			Name:    "vecna_translate_duration_seconds",
+			Help:    "Time spent in the dimension adapter.",
+			Buckets: []float64{0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05},
+		}, []string{"adapter_type"}),
+
+		EndpointPriority: prometheus.NewGaugeVec(prometheus.GaugeOpts{
+			Name: "vecna_endpoint_priority",
+			Help: "Current dynamic routing priority for a forwarding endpoint.",
+		}, []string{"target", "url"}),
+
+		EndpointInflight: prometheus.NewGaugeVec(prometheus.GaugeOpts{
+			Name: "vecna_endpoint_inflight",
+			Help: "Number of active in-flight requests per forwarding endpoint.",
+		}, []string{"target", "url"}),
+
+		EndpointErrorsTotal: prometheus.NewCounterVec(prometheus.CounterOpts{
+			Name: "vecna_endpoint_errors_total",
+			Help: "Total forwarding errors per endpoint, labelled by error type.",
+		}, []string{"target", "url", "error"}),
+
+		TokensTotal: prometheus.NewCounterVec(prometheus.CounterOpts{
+			Name: "vecna_tokens_total",
+			Help: "Tokens consumed by the backing embedding model, by target, model, and token type.",
+		}, []string{"target", "model", "token_type"}),
+	}
+
+	reg.MustRegister(
+		r.RequestsTotal,
+		r.RequestDuration,
+		r.ForwardDuration,
+		r.TranslateDuration,
+		r.EndpointPriority,
+		r.EndpointInflight,
+		r.EndpointErrorsTotal,
+		r.TokensTotal,
+	)
+
+	return r
+}
+
+// Prometheus returns the underlying registry for use with promhttp.HandlerFor.
+func (r *Registry) Prometheus() *prometheus.Registry {
+	return r.reg
+}
+
+// Convenience setters used by the router.
+
+func (r *Registry) SetEndpointPriority(target, url string, v float64) {
+	r.EndpointPriority.WithLabelValues(target, url).Set(v)
+}
+
+func (r *Registry) SetEndpointInflight(target, url string, v float64) {
+	r.EndpointInflight.WithLabelValues(target, url).Set(v)
+}
+
+func (r *Registry) IncEndpointErrors(target, url, errType string) {
+	r.EndpointErrorsTotal.WithLabelValues(target, url, errType).Inc()
+}
+
+func (r *Registry) AddTokens(target, model string, promptTokens, totalTokens int) {
+	if promptTokens > 0 {
+		r.TokensTotal.WithLabelValues(target, model, "prompt").Add(float64(promptTokens))
+	}
+	if totalTokens > 0 {
+		r.TokensTotal.WithLabelValues(target, model, "total").Add(float64(totalTokens))
+	}
+}

pkg/metrics/middleware.go

@@ -0,0 +1,62 @@
+package metrics
+
+import (
+	"fmt"
+	"net/http"
+
+	"github.com/uptrace/bunrouter"
+)
+
+// Middleware returns a bunrouter middleware that records per-request Prometheus metrics.
+// It reads timing from the RequestTrace stored in the context (set by server/trace.go).
+// The trace target/url labels are optional; pass empty strings if not applicable.
+func (r *Registry) Middleware(getTrace func(req bunrouter.Request) TraceSnapshot) bunrouter.MiddlewareFunc {
+	return func(next bunrouter.HandlerFunc) bunrouter.HandlerFunc {
+		return func(w http.ResponseWriter, req bunrouter.Request) error {
+			rw := &statusWriter{ResponseWriter: w, status: http.StatusOK}
+			err := next(rw, req)
+
+			snap := getTrace(req)
+			endpoint := req.URL.Path
+			status := fmt.Sprintf("%d", rw.status)
+
+			r.RequestsTotal.WithLabelValues(endpoint, status).Inc()
+			r.RequestDuration.WithLabelValues(endpoint).Observe(snap.TotalSeconds)
+			if snap.ForwardTarget != "" {
+				r.ForwardDuration.WithLabelValues(snap.ForwardTarget, snap.ForwardURL).Observe(snap.ForwardSeconds)
+			}
+			if snap.AdapterType != "" {
+				r.TranslateDuration.WithLabelValues(snap.AdapterType).Observe(snap.TranslateSeconds)
+			}
+			if snap.PromptTokens > 0 || snap.TotalTokens > 0 {
+				r.AddTokens(snap.ForwardTarget, snap.ForwardModel, snap.PromptTokens, snap.TotalTokens)
+			}
+
+			return err
+		}
+	}
+}
+
+// TraceSnapshot carries the timing and usage values the metrics middleware needs.
+type TraceSnapshot struct {
+	TotalSeconds     float64
+	ForwardSeconds   float64
+	TranslateSeconds float64
+	ForwardTarget    string
+	ForwardURL       string
+	ForwardModel     string
+	AdapterType      string
+	PromptTokens     int
+	TotalTokens      int
+}
+
+// statusWriter wraps http.ResponseWriter to capture the written status code.
+type statusWriter struct {
+	http.ResponseWriter
+	status int
+}
+
+func (sw *statusWriter) WriteHeader(code int) {
+	sw.status = code
+	sw.ResponseWriter.WriteHeader(code)
+}