优化上下文

backend/my_work/models/configs.go

@@ -34,6 +34,7 @@ type ConfigsAIChatOpenAI_ struct {
 	Model               string `mapstructure:"model"`
 	Timeout             int    `mapstructure:"timeout"`
 	MaxTokens           int    `mapstructure:"maxTokens"`
+	ContextWindowTokens int    `mapstructure:"contextWindowTokens"`
 	SystemPrompt        string `mapstructure:"systemPrompt"`
 }
 

backend/my_work/routers/apiAIChat.go

@@ -186,6 +186,7 @@ func handleOpenAIProfiles(ctx *gin.Context) {
 			"model":               profile.Model,
 			"timeout":             profile.Timeout,
 			"maxTokens":           profile.MaxTokens,
+			"contextWindowTokens": profile.ContextWindowTokens,
 		})
 	}
 	ReturnJson(ctx, "apiOK", gin.H{
@@ -317,6 +318,17 @@ func handleChat(ctx *gin.Context) {
 		apiReq.Messages = append([]openaiMessage{{Role: "system", Content: profile.SystemPrompt}}, apiReq.Messages...)
 	}
 
+	trimmedMessages, trimStats := trimOpenAIMessagesToContextWindow(apiReq.Messages, profile.ContextWindowTokens)
+	apiReq.Messages = trimmedMessages
+	if trimStats.RemovedMessages > 0 {
+		emitTrace("model", "context_window", "success", "上下文窗口已裁剪旧消息", map[string]interface{}{
+			"limit":            trimStats.Limit,
+			"before_tokens":    trimStats.BeforeTokens,
+			"after_tokens":     trimStats.AfterTokens,
+			"removed_messages": trimStats.RemovedMessages,
+		})
+	}
+
 	modelPromptTokens := estimateOpenAIMessagesTokens(apiReq.Messages)
 	completionTokens := 0
 	modelUsageReceived := false
@@ -603,6 +615,74 @@ func normalizeImageDataURI(raw string) (string, error) {
 	return "data:" + mimeType + ";base64," + payload, nil
 }
 
+type contextWindowTrimStats struct {
+	Enabled         bool
+	Limit           int
+	BeforeTokens    int
+	AfterTokens     int
+	RemovedMessages int
+}
+
+func trimOpenAIMessagesToContextWindow(messages []openaiMessage, maxTokens int) ([]openaiMessage, contextWindowTrimStats) {
+	stats := contextWindowTrimStats{Enabled: maxTokens > 0, Limit: maxTokens}
+	if maxTokens <= 0 || len(messages) == 0 {
+		stats.BeforeTokens = estimateOpenAIMessagesTokens(messages)
+		stats.AfterTokens = stats.BeforeTokens
+		return messages, stats
+	}
+
+	result := append([]openaiMessage(nil), messages...)
+	stats.BeforeTokens = estimateOpenAIMessagesTokens(result)
+	stats.AfterTokens = stats.BeforeTokens
+	if stats.BeforeTokens <= maxTokens {
+		return result, stats
+	}
+
+	for stats.AfterTokens > maxTokens {
+		startIndex := 0
+		if len(result) > 0 && result[0].Role == "system" {
+			startIndex = 1
+		}
+		latestUserIndex := latestUserMessageIndex(result)
+		removeIndex := -1
+		for i := startIndex; i < len(result); i++ {
+			if i == latestUserIndex {
+				continue
+			}
+			if result[i].Role == "system" {
+				continue
+			}
+			removeIndex = i
+			break
+		}
+		if removeIndex == -1 {
+			break
+		}
+
+		removeCount := 1
+		if result[removeIndex].Role == "user" {
+			nextIndex := removeIndex + 1
+			if nextIndex < len(result) && nextIndex != latestUserIndex && result[nextIndex].Role == "assistant" {
+				removeCount = 2
+			}
+		}
+		result = append(result[:removeIndex], result[removeIndex+removeCount:]...)
+		stats.RemovedMessages += removeCount
+		stats.AfterTokens = estimateOpenAIMessagesTokens(result)
+	}
+
+	return result, stats
+}
+
+func latestUserMessageIndex(messages []openaiMessage) int {
+	for i := len(messages) - 1; i >= 0; i-- {
+		if messages[i].Role == "user" {
+			return i
+		}
+	}
+	return -1
+}
+
 type tokenUsageTracker struct {
 	promptTokens         int
 	completionTokens     int

backend/my_work/routers/apiAIChatConfig.go

@@ -26,6 +26,7 @@ type TabAIChatOpenAIProfile struct {
 	Model               string     `gorm:"size:200"`
 	Timeout             int        `gorm:"default:120"`
 	MaxTokens           int        `gorm:"default:4096"`
+	ContextWindowTokens int        `gorm:"default:0"`
 	SystemPrompt        string     `gorm:"type:text"`
 	SortOrder           int        `gorm:"default:0;index"`
 	CreatedAt           *time.Time `gorm:"type:datetime;autoCreateTime"`
@@ -125,6 +126,7 @@ func seedAIChatConfigFromYAMLIfEmpty() error {
 				BaseUrl:             "https://ark.cn-beijing.volces.com/api/v3",
 				Timeout:             120,
 				MaxTokens:           4096,
+				ContextWindowTokens: 0,
 				SystemPrompt:        "你是一个有帮助的 AI 助手。",
 			}}
 		}
@@ -146,6 +148,7 @@ func seedAIChatConfigFromYAMLIfEmpty() error {
 				Model:               profile.Model,
 				Timeout:             profile.Timeout,
 				MaxTokens:           profile.MaxTokens,
+				ContextWindowTokens: nonNegativeInt(profile.ContextWindowTokens),
 				SystemPrompt:        profile.SystemPrompt,
 				SortOrder:           i,
 			}).Error; err != nil {
@@ -239,6 +242,7 @@ func RefreshAIChatConfigCache() error {
 			Model:               profile.Model,
 			Timeout:             defaultInt(profile.Timeout, 120),
 			MaxTokens:           defaultInt(profile.MaxTokens, 4096),
+			ContextWindowTokens: nonNegativeInt(profile.ContextWindowTokens),
 			SystemPrompt:        profile.SystemPrompt,
 		})
 	}
@@ -269,6 +273,13 @@ func defaultInt(value int, fallback int) int {
 	return value
 }
 
+func nonNegativeInt(value int) int {
+	if value < 0 {
+		return 0
+	}
+	return value
+}
+
 func handleAIChatAdminGetConfig(ctx *gin.Context) {
 	if ok, _ := requireSysAdmin(ctx); !ok {
 		return
@@ -357,6 +368,7 @@ func saveAIChatConfig(req models.ConfigsAIChat_) error {
 			if profile.MaxTokens <= 0 {
 				profile.MaxTokens = 4096
 			}
+			profile.ContextWindowTokens = nonNegativeInt(profile.ContextWindowTokens)
 			if profile.Active {
 				if activeSet {
 					profile.Active = false
@@ -373,6 +385,7 @@ func saveAIChatConfig(req models.ConfigsAIChat_) error {
 				Model:               profile.Model,
 				Timeout:             profile.Timeout,
 				MaxTokens:           profile.MaxTokens,
+				ContextWindowTokens: profile.ContextWindowTokens,
 				SystemPrompt:        profile.SystemPrompt,
 				SortOrder:           i,
 			}
@@ -454,6 +467,7 @@ func maskAIChatProfiles(profiles []models.ConfigsAIChatOpenAI_) []gin.H {
 			"model":               profile.Model,
 			"timeout":             profile.Timeout,
 			"maxTokens":           profile.MaxTokens,
+			"contextWindowTokens": profile.ContextWindowTokens,
 			"systemPrompt":        profile.SystemPrompt,
 		})
 	}

frontend/ops_vue_js/src/i18n/en.json

@@ -113,6 +113,8 @@
     "model": "Model",
     "timeout": "Timeout (seconds)",
     "max_tokens": "Max Tokens",
+    "context_window_tokens": "Context Window Tokens",
+    "context_window_tokens_hint": "0 means unlimited. When exceeded, the system prompt and latest messages are kept while oldest Q/A turns are removed.",
     "system_prompt": "System Prompt",
     "tool_router": "Tool Router",
     "router_profile": "Router Profile",
@@ -132,6 +134,7 @@
     "error_api_key_required": "When AI is enabled, the default profile must have an API key",
     "error_timeout": "Timeout must be a positive integer",
     "error_max_tokens": "Max tokens must be a positive integer",
+    "error_context_window_tokens": "Context window tokens must be a non-negative integer",
     "error_router_profile": "Tool router profile must exist in profile list",
     "error_tool_name_required": "Tool name is required",
     "error_tool_name_duplicate": "Tool names must be unique"

frontend/ops_vue_js/src/i18n/zh-CN.json

@@ -113,6 +113,8 @@
     "model": "模型",
     "timeout": "超时（秒）",
     "max_tokens": "最大 Token",
+    "context_window_tokens": "上下文窗口 Token",
+    "context_window_tokens_hint": "0 表示不限制；超过后会保留系统提示词和最新消息，删除最早的问答记录。",
     "system_prompt": "系统提示词",
     "tool_router": "工具路由",
     "router_profile": "路由接口",
@@ -132,6 +134,7 @@
     "error_api_key_required": "启用 AI 时默认接口必须配置 API Key",
     "error_timeout": "超时必须是正整数",
     "error_max_tokens": "最大 Token 必须是正整数",
+    "error_context_window_tokens": "上下文窗口 Token 必须是非负整数",
     "error_router_profile": "工具路由接口必须来自现有接口列表",
     "error_tool_name_required": "工具名称不能为空",
     "error_tool_name_duplicate": "工具名称不能重复"

frontend/ops_vue_js/src/views/admin/AIConfigView.vue

@@ -42,6 +42,7 @@ function normalizeProfile(profile = {}) {
     model: profile.model || '',
     timeout: Number(profile.timeout || 120),
     maxTokens: Number(profile.maxTokens || 4096),
+    contextWindowTokens: Number(profile.contextWindowTokens || 0),
     systemPrompt: profile.systemPrompt || '',
   }
 }
@@ -117,6 +118,7 @@ function addProfile() {
     model: '',
     timeout: 120,
     maxTokens: 4096,
+    contextWindowTokens: 0,
     systemPrompt: '',
   })
 }
@@ -178,6 +180,7 @@ function validate() {
 
     profile.timeout = Number(profile.timeout)
     profile.maxTokens = Number(profile.maxTokens)
+    profile.contextWindowTokens = Number(profile.contextWindowTokens)
     if (!Number.isInteger(profile.timeout) || profile.timeout <= 0) {
       toast.error(t('aiconfig.error_timeout'))
       return false
@@ -186,6 +189,10 @@ function validate() {
       toast.error(t('aiconfig.error_max_tokens'))
       return false
     }
+    if (!Number.isInteger(profile.contextWindowTokens) || profile.contextWindowTokens < 0) {
+      toast.error(t('aiconfig.error_context_window_tokens'))
+      return false
+    }
   }
 
   if (!hasActive) {
@@ -248,6 +255,7 @@ function buildPayload() {
       model: profile.model.trim(),
       timeout: Number(profile.timeout),
       maxTokens: Number(profile.maxTokens),
+      contextWindowTokens: Number(profile.contextWindowTokens),
       systemPrompt: profile.systemPrompt || '',
     })),
     toolRouter: {
@@ -360,6 +368,11 @@ async function saveConfig() {
               <span>{{ t('aiconfig.max_tokens') }}</span>
               <input v-model.number="profile.maxTokens" class="input" type="number" min="1" />
             </label>
+            <label class="field">
+              <span>{{ t('aiconfig.context_window_tokens') }}</span>
+              <input v-model.number="profile.contextWindowTokens" class="input" type="number" min="0" />
+              <span class="text-xs font-normal text-gray-500 dark:text-dk-subtle">{{ t('aiconfig.context_window_tokens_hint') }}</span>
+            </label>
             <label class="field md:col-span-2">
               <span>{{ t('aiconfig.system_prompt') }}</span>
               <textarea v-model="profile.systemPrompt" class="input min-h-24 resize-y" />