Losita/smartmate
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Version: 0.5.0.dev.260310

Browse Source 
refactor: ♻️ 调整 Outbox 写入时序并移除 Kafka 首包同步投递逻辑

* 将 `outbox` 表写入逻辑后置到 LLM 请求之后，减少主链路阻塞
* 删除 Codex 生成的 Kafka 首包同步投递抽象逻辑，简化消息发送流程
* 优化 SSE 首字到达时间，整体降低约 1s 延迟
* 当前在请求 LLM 之前的流程全部为 Redis 操作，显著降低 IO 开销

docs: 📊 保留 SSE 链路性能打点逻辑

* 保留原有 SSE 全链路打点计时代码，便于后续性能排查与分析
* 当前默认注释，如需使用可手动启用进行性能调试
This commit is contained in:
Losita
2026-03-10 23:10:09 +08:00
parent 959049db42
commit 912a6d8cfe
6 changed files with 138 additions and 123 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
backend/agent/graph.go
View File

@@ -13,7 +13,7 @@ import (
arkModel "github.com/volcengine/volcengine-go-sdk/service/arkruntime/model"
)
// StreamResponse 为 OpenAI/DeepSeek 兼容的流式 chunk 结构。
// StreamResponse 是 OpenAI/DeepSeek 兼容的流式 chunk 结构。
type StreamResponse struct {
ID string `json:"id"`
Object string `json:"object"`
@@ -88,8 +88,24 @@ func ToOpenAIFinishStream(requestID, modelName string, created int64) (string, e
return string(jsonBytes), nil
}
func StreamChat(ctx context.Context, llm *ark.ChatModel, modelName string, userInput string, ifThinking bool, chatHistory []*schema.Message, outChan chan<- string) (string, error) {
// 1) 组装提示消息
// StreamChat 负责模型流式输出，并在关键节点打点：
// 1) 流连接建立（llm.Stream 返回）
// 2) 首包到达（首字延迟）
// 3) 流式输出结束
func StreamChat(
ctx context.Context,
llm *ark.ChatModel,
modelName string,
userInput string,
ifThinking bool,
chatHistory []*schema.Message,
outChan chan<- string,
traceID string,
chatID string,
requestStart time.Time,
) (string, error) {
/*callStart := time.Now()*/
messages := make([]*schema.Message, 0)
messages = append(messages, schema.SystemMessage(SystemPrompt))
if len(chatHistory) > 0 {
@@ -97,13 +113,14 @@ func StreamChat(ctx context.Context, llm *ark.ChatModel, modelName string, userI
}
messages = append(messages, schema.UserMessage(userInput))
// 2) 发起流式请求
var thinking *ark.Thinking
if ifThinking {
thinking = &arkModel.Thinking{Type: arkModel.ThinkingTypeEnabled}
} else {
thinking = &arkModel.Thinking{Type: arkModel.ThinkingTypeDisabled}
}
/*connectStart := time.Now()*/
reader, err := llm.Stream(ctx, messages, ark.WithThinking(thinking))
if err != nil {
return "", err
@@ -116,8 +133,18 @@ func StreamChat(ctx context.Context, llm *ark.ChatModel, modelName string, userI
requestID := "chatcmpl-" + uuid.NewString()
created := time.Now().Unix()
firstChunk := true
chunkCount := 0
/*streamRecvStart := time.Now()
log.Printf("打点|流连接建立|trace_id=%s|chat_id=%s|request_id=%s|本步耗时_ms=%d|请求累计_ms=%d|history_len=%d",
traceID,
chatID,
requestID,
time.Since(connectStart).Milliseconds(),
time.Since(requestStart).Milliseconds(),
len(chatHistory),
)*/
// 3) 持续转发 chunk
var fullText strings.Builder
for {
chunk, err := reader.Recv()
@@ -136,11 +163,20 @@ func StreamChat(ctx context.Context, llm *ark.ChatModel, modelName string, userI
}
if payload != "" {
outChan <- payload
firstChunk = false
chunkCount++
/*if firstChunk {
log.Printf("打点|首包到达|trace_id=%s|chat_id=%s|request_id=%s|本步耗时_ms=%d|请求累计_ms=%d",
traceID,
chatID,
requestID,
time.Since(streamRecvStart).Milliseconds(),
time.Since(requestStart).Milliseconds(),
)
firstChunk = false
}*/
}
}
// 4) 发送结束 chunk 和 [DONE]
finishChunk, err := ToOpenAIFinishStream(requestID, modelName, created)
if err != nil {
return "", err
@@ -148,5 +184,15 @@ func StreamChat(ctx context.Context, llm *ark.ChatModel, modelName string, userI
outChan <- finishChunk
outChan <- "[DONE]"
/*log.Printf("打点|流式输出结束|trace_id=%s|chat_id=%s|request_id=%s|chunks=%d|reply_chars=%d|本步耗时_ms=%d|请求累计_ms=%d",
traceID,
chatID,
requestID,
chunkCount,
len(fullText.String()),
time.Since(callStart).Milliseconds(),
time.Since(requestStart).Milliseconds(),
)*/
return fullText.String(), nil
}