Bosch_Demo/LaodingBot
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: optimize WebUI stream output and sanitize user-facing answers

Browse Source
This commit is contained in:
Ding, Shuo
2026-03-13 13:14:37 +08:00
parent 8dc5354fa4
commit 33c357a1de
8 changed files with 228 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

86
internal/agent/orchestrator.go
View File

@@ -358,16 +358,17 @@ func (o *Orchestrator) buildUnifiedSystemPrompt(userInput string, routedSkills [
"",
"===== ReAct 思考指引 =====",
"你采用 ReAct（Reasoning + Acting）模式进行任务处理。",
"1. 思考优先：在做出任何行动之前，先在回复中阐述你的推理过程（Thought）。",
"1. 思考优先：在做出任何行动之前，先完成内部推理，但不要把 Thought、trace、step 暴露给用户。",
"2. 工具调用：如果需要获取信息或执行操作，使用提供的工具函数（function calling）进行调用。",
"3. 观察反馈：检查工具返回的结果，据此决定下一步行动。",
"4. 最终回答：当你有足够信息时，直接给出面向用户的最终文本回复，不要调用工具。",
"4. 最终回答：当你有足够信息时，只输出面向用户的最终文本回复，不要附带推理轨迹，不要调用工具。",
"",
"注意事项：",
"- 每次要么调用工具，要么给出最终回答，不要两者都做。",
"- 如果工具调用失败，根据错误信息（Traceback）调整策略后重试或给出替代方案。",
"- 涉及文件、目录、命令时，优先调用工具获取真实结果，不要猜测。",
"- 你的思考过程（Thought）应写在回复内容中，帮助追踪推理逻辑。",
"- 如果本轮需要调用工具，可以在 assistant content 中写简短内部推理，供系统记录日志；这些内容不会直接展示给用户。",
"- 最终用户可见内容中禁止出现 Thought、Trace、Step、Observation、Action、ActionInput 等字段或标题。",
"",
"===== 运行环境 =====",
runtimeDoc,
@@ -556,7 +557,7 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
const maxSteps = 20
for step := 1; step <= maxSteps; step++ {
if o.log != nil {
o.log.Infof("%s react step=%d start messages_count=%d", traceLogPrefix, step, len(messages))
o.log.Debugf("%s react step=%d start messages_count=%d", traceLogPrefix, step, len(messages))
}
// 调用 LLM（传入完整 messages + tools 定义）
@@ -566,7 +567,7 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
}
if o.log != nil {
o.log.Infof("%s react step=%d content_len=%d tool_calls=%d",
o.log.Debugf("%s react step=%d content_len=%d tool_calls=%d",
traceLogPrefix, step, len(completion.Content), len(completion.ToolCalls))
if completion.Content != "" {
o.log.Debugf("%s react step=%d thought=%q", traceLogPrefix, step, completion.Content)
@@ -575,12 +576,12 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
// ========== 无 tool_calls → 最终回答 ==========
if len(completion.ToolCalls) == 0 {
finalText := strings.TrimSpace(completion.Content)
finalText := sanitizeUserFacingAnswer(completion.Content)
if finalText == "" {
finalText = "已完成处理。"
}
if o.log != nil {
o.log.Infof("%s react final at step=%d answer_len=%d", traceLogPrefix, step, len(finalText))
o.log.Debugf("%s react final at step=%d answer_len=%d", traceLogPrefix, step, len(finalText))
}
return finalText, nil
}
@@ -614,7 +615,7 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
}
if o.log != nil {
o.log.Infof("%s react step=%d tool_call tool=%s input_len=%d", traceLogPrefix, step, toolName, len(toolInput))
o.log.Debugf("%s react step=%d tool_call tool=%s input_len=%d", traceLogPrefix, step, toolName, len(toolInput))
o.log.Debugf("%s react step=%d tool=%s input=%q", traceLogPrefix, step, toolName, toolInput)
}
@@ -633,7 +634,7 @@ func (o *Orchestrator) runUnifiedReAct(ctx context.Context, chatID, userID, comp
}
if o.log != nil {
o.log.Infof("%s react step=%d tool=%s observation_len=%d", traceLogPrefix, step, toolName, len(obs))
o.log.Debugf("%s react step=%d tool=%s observation_len=%d", traceLogPrefix, step, toolName, len(obs))
o.log.Debugf("%s react step=%d tool=%s observation=%q", traceLogPrefix, step, toolName, truncateForLog(obs, 500))
}
@@ -711,7 +712,7 @@ func (o *Orchestrator) runUnifiedReActStream(ctx context.Context, chatID, userID
const maxSteps = 20
for step := 1; step <= maxSteps; step++ {
if o.log != nil {
o.log.Infof("%s react stream step=%d start messages_count=%d", traceLogPrefix, step, len(messages))
o.log.Debugf("%s react stream step=%d start messages_count=%d", traceLogPrefix, step, len(messages))
}
// 调用 LLM
@@ -721,7 +722,7 @@ func (o *Orchestrator) runUnifiedReActStream(ctx context.Context, chatID, userID
}
if o.log != nil {
o.log.Infof("%s react stream step=%d content_len=%d tool_calls=%d",
o.log.Debugf("%s react stream step=%d content_len=%d tool_calls=%d",
traceLogPrefix, step, len(completion.Content), len(completion.ToolCalls))
if completion.Content != "" {
o.log.Debugf("%s react stream step=%d thought=%q", traceLogPrefix, step, completion.Content)
@@ -741,12 +742,12 @@ func (o *Orchestrator) runUnifiedReActStream(ctx context.Context, chatID, userID
// ========== 无 tool_calls → 最终回答 ==========
if len(completion.ToolCalls) == 0 {
finalText := strings.TrimSpace(completion.Content)
finalText := sanitizeUserFacingAnswer(completion.Content)
if finalText == "" {
finalText = "已完成处理。"
}
if o.log != nil {
o.log.Infof("%s react stream final at step=%d answer_len=%d", traceLogPrefix, step, len(finalText))
o.log.Debugf("%s react stream final at step=%d answer_len=%d", traceLogPrefix, step, len(finalText))
}
// 推送最终答案事件
if err := callback(StreamEvent{
@@ -808,7 +809,7 @@ func (o *Orchestrator) runUnifiedReActStream(ctx context.Context, chatID, userID
}
if o.log != nil {
o.log.Infof("%s react stream step=%d tool_call tool=%s input_len=%d", traceLogPrefix, step, toolName, len(toolInput))
o.log.Debugf("%s react stream step=%d tool_call tool=%s input_len=%d", traceLogPrefix, step, toolName, len(toolInput))
o.log.Debugf("%s react stream step=%d tool=%s input=%q", traceLogPrefix, step, toolName, toolInput)
}
@@ -827,7 +828,7 @@ func (o *Orchestrator) runUnifiedReActStream(ctx context.Context, chatID, userID
}
if o.log != nil {
o.log.Infof("%s react stream step=%d tool=%s observation_len=%d", traceLogPrefix, step, toolName, len(obs))
o.log.Debugf("%s react stream step=%d tool=%s observation_len=%d", traceLogPrefix, step, toolName, len(obs))
o.log.Debugf("%s react stream step=%d tool=%s observation=%q", traceLogPrefix, step, toolName, truncateForLog(obs, 500))
}
@@ -872,7 +873,7 @@ func (o *Orchestrator) runLegacyReAct(ctx context.Context, chatID, userID, compr
for step := 1; step <= maxSteps; step++ {
if o.log != nil {
o.log.Infof("%s legacy react step=%d start", traceLogPrefix, step)
o.log.Debugf("%s legacy react step=%d start", traceLogPrefix, step)
}
messages := buildReActMessages(systemPrompt, compressedContext, userInput, scratchpad)
@@ -890,10 +891,10 @@ func (o *Orchestrator) runLegacyReAct(ctx context.Context, chatID, userID, compr
if decision.IsFinalAnswer {
finalText := ""
if decision.FinalAnswer != nil {
finalText = strings.TrimSpace(*decision.FinalAnswer)
finalText = sanitizeUserFacingAnswer(*decision.FinalAnswer)
}
if finalText == "" {
finalText = strings.TrimSpace(decision.Thought)
finalText = sanitizeUserFacingAnswer(decision.Thought)
}
if finalText == "" {
finalText = "已完成处理。"
@@ -1518,3 +1519,52 @@ func truncateForLog(s string, maxLen int) string {
}
return s[:maxLen] + "...(truncated)"
}
func sanitizeUserFacingAnswer(raw string) string {
raw = strings.ReplaceAll(raw, "\r\n", "\n")
raw = strings.TrimSpace(raw)
if raw == "" {
return ""
}
markers := []string{"Final Answer:", "Final Answer：", "最终回答:", "最终回答：", "最终答案:", "最终答案：", "Answer:", "Answer："}
for _, marker := range markers {
idx := strings.LastIndex(raw, marker)
if idx >= 0 {
candidate := strings.TrimSpace(raw[idx+len(marker):])
if candidate != "" {
return candidate
}
}
}
lines := strings.Split(raw, "\n")
cleaned := make([]string, 0, len(lines))
for _, line := range lines {
trimmed := strings.TrimSpace(line)
if trimmed == "" {
if len(cleaned) > 0 && cleaned[len(cleaned)-1] != "" {
cleaned = append(cleaned, "")
}
continue
}
lower := strings.ToLower(trimmed)
if strings.HasPrefix(lower, "thought:") || strings.HasPrefix(lower, "trace:") || strings.HasPrefix(lower, "observation:") ||
strings.HasPrefix(lower, "action:") || strings.HasPrefix(lower, "actioninput:") || strings.HasPrefix(lower, "action input:") ||
strings.HasPrefix(lower, "step ") || strings.HasPrefix(trimmed, "思考：") || strings.HasPrefix(trimmed, "思考:") ||
strings.HasPrefix(trimmed, "推理：") || strings.HasPrefix(trimmed, "推理:") || strings.HasPrefix(trimmed, "观察：") ||
strings.HasPrefix(trimmed, "观察:") || strings.HasPrefix(trimmed, "行动：") || strings.HasPrefix(trimmed, "行动:") ||
strings.HasPrefix(trimmed, "步骤 ") {
continue
}
cleaned = append(cleaned, trimmed)
}
for len(cleaned) > 0 && cleaned[len(cleaned)-1] == "" {
cleaned = cleaned[:len(cleaned)-1]
}
if len(cleaned) == 0 {
return ""
}
return strings.TrimSpace(strings.Join(cleaned, "\n"))
}

29
internal/agent/orchestrator_skill_selection_test.go
View File

@@ -93,3 +93,32 @@ func TestMatchSkillsByNameEmpty(t *testing.T) {
t.Fatalf("expected 0 matches, got %d", len(matched))
}
}
func TestSanitizeUserFacingAnswerExtractsFinalAnswer(t *testing.T) {
raw := "Thought: 先分析用户问题\nObservation: 已经有足够信息\nFinal Answer: 这是给用户的结果"
got := sanitizeUserFacingAnswer(raw)
if got != "这是给用户的结果" {
t.Fatalf("expected final answer only, got %q", got)
}
}
func TestSanitizeUserFacingAnswerDropsTraceLines(t *testing.T) {
raw := strings.Join([]string{
"Step 1 Thought: 检查上下文",
"Action: shell",
"Observation: ok",
"请执行以下变更。",
}, "\n")
got := sanitizeUserFacingAnswer(raw)
if got != "请执行以下变更。" {
t.Fatalf("expected user-facing text only, got %q", got)
}
}
func TestSanitizeUserFacingAnswerKeepsNormalAnswer(t *testing.T) {
raw := "1. 先打开配置文件\n2. 修改端口后重启服务"
got := sanitizeUserFacingAnswer(raw)
if got != raw {
t.Fatalf("expected answer unchanged, got %q", got)
}
}

33
internal/llm/client.go
View File

@@ -81,9 +81,10 @@ type InputFile struct {
}
type OpenAICompatibleClient struct {
client openai.Client
model string
log *logger.Logger
client openai.Client
model string
disableThinkingParam bool
log *logger.Logger
}
func NewOpenAICompatibleClient(cfg config.LLMConfig, log *logger.Logger) *OpenAICompatibleClient {
@@ -95,9 +96,10 @@ func NewOpenAICompatibleClient(cfg config.LLMConfig, log *logger.Logger) *OpenAI
opts = append(opts, option.WithBaseURL(cfg.BaseURL))
}
return &OpenAICompatibleClient{
client: openai.NewClient(opts...),
model: cfg.Model,
log: log,
client: openai.NewClient(opts...),
model: cfg.Model,
disableThinkingParam: shouldDisableThinkingParam(cfg.BaseURL),
log: log,
}
}
@@ -138,7 +140,7 @@ func (c *OpenAICompatibleClient) GenerateWithTools(ctx context.Context, messages
}
}
resp, err := c.client.Chat.Completions.New(ctx, params)
resp, err := c.client.Chat.Completions.New(ctx, params, c.chatCompletionRequestOptions()...)
if err != nil {
return nil, fmt.Errorf("llm tool-call request failed: %w", err)
}
@@ -180,7 +182,7 @@ func (c *OpenAICompatibleClient) generateWithMessagesInternal(ctx context.Contex
Messages: sdkMessages,
}
resp, err := c.client.Chat.Completions.New(ctx, params)
resp, err := c.client.Chat.Completions.New(ctx, params, c.chatCompletionRequestOptions()...)
if err != nil {
if c.log != nil {
c.log.Errorf("llm request failed err=%v", err)
@@ -392,3 +394,18 @@ func appendIfMissing(items []string, value string) []string {
}
return append(items, value)
}
func (c *OpenAICompatibleClient) chatCompletionRequestOptions() []option.RequestOption {
if !c.disableThinkingParam {
return nil
}
return []option.RequestOption{option.WithJSONSet("enable_thinking", false)}
}
func shouldDisableThinkingParam(baseURL string) bool {
baseURL = strings.ToLower(strings.TrimSpace(baseURL))
if baseURL == "" {
return false
}
return strings.Contains(baseURL, "dashscope.aliyuncs.com")
}

24
internal/llm/client_test.go Normal file
View File

@@ -0,0 +1,24 @@
package llm
import "testing"
func TestShouldDisableThinkingParam(t *testing.T) {
if !shouldDisableThinkingParam("https://dashscope.aliyuncs.com/compatible-mode/v1") {
t.Fatal("expected DashScope base URL to require enable_thinking=false")
}
if shouldDisableThinkingParam("https://api.openai.com/v1") {
t.Fatal("expected standard OpenAI base URL not to require enable_thinking=false")
}
}
func TestChatCompletionRequestOptions(t *testing.T) {
client := &OpenAICompatibleClient{disableThinkingParam: true}
if got := len(client.chatCompletionRequestOptions()); got != 1 {
t.Fatalf("expected 1 request option when disableThinkingParam=true, got %d", got)
}
client.disableThinkingParam = false
if got := len(client.chatCompletionRequestOptions()); got != 0 {
t.Fatalf("expected 0 request options when disableThinkingParam=false, got %d", got)
}
}