feat: metrics api cost time

server/src/core/fetch.go

@@ -37,7 +37,10 @@ func (pr *row) fetchRemote() (ab []byte, err error) {
 	client := &http.Client{
 		// Timeout: 30 * time.Second,
 	}
+	t := time.Now()
 	rsp, err := client.Do(req)
+	costMs := uint32(time.Since(t) / time.Millisecond)
+	metrics.OpenAITime(costMs)
 	if err != nil {
 		fmt.Fprintf(b, "client.Do fail: %s\n", err.Error())
 		return

server/src/metrics/init.go

@@ -21,4 +21,7 @@ func init() {
 
 	prometheus.MustRegister(limitReq)
 	prometheus.MustRegister(limitToken)
+
+	prometheus.MustRegister(openaiTime)
+	prometheus.MustRegister(vaTime)
 }

server/src/metrics/openai.go

@@ -0,0 +1,10 @@
+package metrics
+
+var (
+	openaiTime = newSummary(`orca_openai_time`, `openai api time`)
+)
+
+// OpenAITime ...
+func OpenAITime(ms uint32) {
+	openaiTime.Observe(float64(ms))
+}

server/src/metrics/va.go

@@ -0,0 +1,10 @@
+package metrics
+
+var (
+	vaTime = newSummary(`orca_va_time`, `vertexai api time`)
+)
+
+// VaTime ...
+func VaTime(ms uint32) {
+	vaTime.Observe(float64(ms))
+}

server/src/vertexai/chat.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"project/metrics"
 	"project/pb"
 	"project/util"
 	"time"
@@ -168,6 +169,7 @@ func loadChat(k chatKey) (*ChatRsp, error) {
 		Raw:    rsp,
 		CostMs: uint32(time.Since(t) / time.Millisecond),
 	}
+	metrics.VaTime(r.CostMs)
 
 	answer := &pb.VaChatRsp{}
 	answer.Content, err = getChatVal(rsp)