From aa4af5365ec3c08e28522029053d40b2294910c6 Mon Sep 17 00:00:00 2001 From: wizardchen Date: Tue, 20 Jan 2026 16:20:43 +0800 Subject: [PATCH] feat: add thinking mode support for agents --- frontend/src/i18n/locales/en-US.ts | 1 + frontend/src/i18n/locales/zh-CN.ts | 1 + frontend/src/views/agent/AgentEditorModal.vue | 18 + frontend/src/views/chat/components/botmsg.vue | 20 +- .../src/views/chat/components/deepThink.vue | 229 +++--- frontend/src/views/chat/index.vue | 129 +++- internal/agent/engine.go | 7 +- .../chat_pipline/chat_completion_stream.go | 49 ++ internal/application/service/session.go | 7 + internal/handler/session/helpers.go | 14 +- internal/handler/session/qa.go | 14 +- internal/models/chat/chat.go | 40 +- internal/models/chat/lkeap.go | 77 ++ internal/models/chat/provider_chat.go | 82 ++ internal/models/chat/qwen.go | 62 ++ internal/models/chat/remote_api.go | 716 ++++++++++-------- internal/models/chat/sse_reader.go | 59 ++ internal/models/provider/lkeap.go | 65 ++ internal/models/provider/provider.go | 5 + internal/types/agent.go | 2 + internal/types/custom_agent.go | 2 + 21 files changed, 1118 insertions(+), 481 deletions(-) create mode 100644 internal/models/chat/lkeap.go create mode 100644 internal/models/chat/provider_chat.go create mode 100644 internal/models/chat/qwen.go create mode 100644 internal/models/chat/sse_reader.go create mode 100644 internal/models/provider/lkeap.go diff --git a/frontend/src/i18n/locales/en-US.ts b/frontend/src/i18n/locales/en-US.ts index ea47e6d9..55b5daa9 100755 --- a/frontend/src/i18n/locales/en-US.ts +++ b/frontend/src/i18n/locales/en-US.ts @@ -312,6 +312,7 @@ export default { availablePlaceholders: 'Available Placeholders', placeholderHint: 'Type {{ to trigger autocomplete', temperature: 'Temperature', + thinking: 'Thinking Mode', welcomeMessage: 'Welcome Message', welcomeMessagePlaceholder: 'Message displayed when this agent is selected', suggestedPrompts: 'Suggested Prompts', diff --git a/frontend/src/i18n/locales/zh-CN.ts b/frontend/src/i18n/locales/zh-CN.ts index 712e6aeb..b098eeb6 100755 --- a/frontend/src/i18n/locales/zh-CN.ts +++ b/frontend/src/i18n/locales/zh-CN.ts @@ -842,6 +842,7 @@ export default { availablePlaceholders: "可用占位符", placeholderHint: "输入 {'{{'} 触发自动补全", temperature: "温度", + thinking: "思考模式", welcomeMessage: "欢迎消息", welcomeMessagePlaceholder: "选择该智能体时显示的欢迎消息", suggestedPrompts: "推荐问题", diff --git a/frontend/src/views/agent/AgentEditorModal.vue b/frontend/src/views/agent/AgentEditorModal.vue index d6f7e07a..37c4ce29 100644 --- a/frontend/src/views/agent/AgentEditorModal.vue +++ b/frontend/src/views/agent/AgentEditorModal.vue @@ -307,6 +307,17 @@ + + +
+
+ +

启用模型的扩展思考能力(需要模型支持)

+
+
+ +
+
@@ -1173,6 +1184,7 @@ const defaultFormData = { rerank_model_id: '', temperature: 0.7, max_completion_tokens: 2048, + thinking: false, // 默认禁用思考模式 // Agent模式设置 max_iterations: 10, allowed_tools: [] as string[], @@ -1223,6 +1235,12 @@ const agentMode = computed({ const isAgentMode = computed(() => agentMode.value === 'smart-reasoning'); +// 思考模式计算属性(直接绑定 boolean) +const thinkingEnabled = computed({ + get: () => formData.value.config.thinking === true, + set: (val: boolean) => { formData.value.config.thinking = val; } +}); + // 是否为内置智能体 const isBuiltinAgent = computed(() => { return formData.value.is_builtin === true; diff --git a/frontend/src/views/chat/components/botmsg.vue b/frontend/src/views/chat/components/botmsg.vue index c9bf719e..7353eaf0 100644 --- a/frontend/src/views/chat/components/botmsg.vue +++ b/frontend/src/views/chat/components/botmsg.vue @@ -24,16 +24,9 @@
- -
-
- - - -
-
-
+ +
@@ -136,6 +129,12 @@ const markdownTokens = computed(() => { return marked.lexer(safeMarkdown); }); +// 计算属性:判断是否有实际内容(非空且不只是空白) +const hasActualContent = computed(() => { + const text = props.content || props.session?.content || ''; + return text && text.trim().length > 0; +}); + // 渲染单个 token 为 HTML const renderToken = (token) => { try { @@ -428,8 +427,7 @@ onBeforeUnmount(() => { } .thinking-loading { - margin-left: 16px; - margin-bottom: 8px; + padding: 8px 0; } .loading-typing { diff --git a/frontend/src/views/chat/components/deepThink.vue b/frontend/src/views/chat/components/deepThink.vue index 67745dbf..74194ac8 100644 --- a/frontend/src/views/chat/components/deepThink.vue +++ b/frontend/src/views/chat/components/deepThink.vue @@ -1,38 +1,32 @@ \ No newline at end of file + diff --git a/frontend/src/views/chat/index.vue b/frontend/src/views/chat/index.vue index ebe4bcd8..2226208e 100644 --- a/frontend/src/views/chat/index.vue +++ b/frontend/src/views/chat/index.vue @@ -243,12 +243,21 @@ const handleMsgList = async (data, isScrollType = false, newScrollHeight) => { item.thinkContent = ""; item.content = item.content; item.showThink = false; + item.thinking = false; } else if (item.content.includes('<\/think>')) { + // 历史消息中包含完整的 ... 标签,说明 thinking 已完成 const arr = item.content.trim().split('<\/think>'); item.showThink = true; + item.thinking = false; // 关键:标记 thinking 已完成,使 deepThink 默认折叠 item.thinkContent = arr[0].trim().replace('', ''); let index = item.content.trim().lastIndexOf('<\/think>') item.content = item.content.substring(index + 8); + } else if (item.content.includes('')) { + // 内容包含 但没有 ,说明 thinking 还在进行中(不太可能出现在历史消息中) + item.showThink = true; + item.thinking = true; + item.thinkContent = item.content.replace('', '').trim(); + item.content = ''; } } @@ -408,16 +417,55 @@ onChunk((data) => { } // 判断是否是 Agent 模式的响应 - const isAgentResponse = data.response_type === 'thinking' || - data.response_type === 'tool_call' || - data.response_type === 'tool_result' || - data.response_type === 'references' || - data.response_type === 'answer' || - data.response_type === 'reflection' || - data.response_type === 'stop'; + // 注意:'answer', 'complete', 'references' 类型可能在两种模式下都存在 + // 只有 'thinking', 'tool_call', 'tool_result', 'reflection' 是 Agent 专有的 + const isAgentOnlyResponse = data.response_type === 'thinking' || + data.response_type === 'tool_call' || + data.response_type === 'tool_result' || + data.response_type === 'reflection'; + + // 检查当前消息是否已经是 Agent 模式 + const lastMessage = messagesList[messagesList.length - 1]; + const isCurrentlyAgentMode = lastMessage?.isAgentMode === true; + + // 如果是 Agent 专有的响应类型,或者当前消息已经是 Agent 模式,则走 Agent 处理 + const shouldHandleAsAgent = isAgentOnlyResponse || isCurrentlyAgentMode; + + // 处理 references 事件 - 在两种模式下都需要处理,但不改变模式 + if (data.response_type === 'references') { + // 如果当前是 Agent 模式,走 Agent 处理 + if (isCurrentlyAgentMode) { + handleAgentChunk(data); + return; + } + // 非 Agent 模式:将 references 保存到消息中供 botmsg 使用 + let existingMessage = messagesList.findLast((item) => item.request_id === data.id || item.id === data.id); + + // 如果消息还不存在,先创建一个空的 assistant 消息 + if (!existingMessage) { + existingMessage = { + id: data.id, + request_id: data.id, + role: 'assistant', + content: '', + showThink: false, + thinkContent: '', + thinking: false, + is_completed: false, + knowledge_references: [] + }; + messagesList.push(existingMessage); + loading.value = false; // 消息已创建,关闭 loading + scrollToBottom(); + } + + existingMessage.knowledge_references = data.knowledge_references || data.data?.references || []; + console.log('[References] Saved to message, count:', existingMessage.knowledge_references.length); + return; + } // Agent 模式处理(包括 stop 事件) - if (isAgentResponse || messagesList[messagesList.length - 1]?.isAgentMode) { + if (shouldHandleAsAgent) { // 在 handleAgentChunk 中处理 loading 状态 handleAgentChunk(data); @@ -432,7 +480,23 @@ onChunk((data) => { return; } - // 原有的知识库 QA 处理逻辑 + // 原有的知识库 QA 处理逻辑(非 Agent 模式) + // answer 内容中可能包含 ... 标签 + + // 检查消息是否已经完成,如果已完成则忽略后续的完成事件(防止空内容覆盖) + const existingMessage = messagesList.findLast((item) => { + if (item.request_id === data.id) { + return true + } + return item.id === data.id; + }); + + // 如果消息已完成且当前事件是完成事件(done=true 且无内容),直接忽略 + if (existingMessage?.is_completed && data.done && !data.content) { + console.log('[Non-Agent] Ignoring duplicate completion event for completed message'); + return; + } + fullContent.value += data.content; let obj = { ...data, content: '', role: 'assistant', showThink: false, is_completed: false }; @@ -451,13 +515,6 @@ onChunk((data) => { obj.content = fullContent.value; } - // 检查是否已有消息,如果没有则说明这是第一次,关闭 loading - const existingMessage = messagesList.findLast((item) => { - if (item.request_id === obj.id) { - return true - } - return item.id === obj.id; - }); if (!existingMessage) { loading.value = false; // 消息即将创建,关闭 loading } @@ -704,8 +761,20 @@ const handleAgentChunk = (data) => { case 'answer': // 最终答案 message.thinking = false; - message.content = (message.content || '') + (data.content || ''); - fullContent.value += data.content || ''; + + console.log('[Answer Event] Received:', { + has_content: !!data.content, + content_length: data.content?.length || 0, + done: data.done, + current_message_content_length: message.content?.length || 0 + }); + + // 只有当有实际内容时才追加,避免空内容覆盖 + if (data.content) { + message.content = (message.content || '') + data.content; + fullContent.value += data.content; + console.log('[Answer] Content appended, new length:', message.content.length); + } // Add or update answer event in agentEventStream if (!message.agentEventStream) message.agentEventStream = []; @@ -718,13 +787,19 @@ const handleAgentChunk = (data) => { done: false }; message.agentEventStream.push(answerEvent); + console.log('[Answer] Created new answer event in stream'); } - answerEvent.content = message.content; - answerEvent.done = data.done; + // 只有当有实际内容时才更新 answerEvent.content + if (data.content) { + answerEvent.content = message.content; + console.log('[Answer] answerEvent.content updated, length:', answerEvent.content.length); + } - if (data.done) { - console.log('[Agent] Answer done, content length:', message.content?.length || 0); + // 只在第一次收到 done:true 时标记完成,忽略后续重复的完成事件 + if (data.done && !answerEvent.done) { + answerEvent.done = true; + console.log('[Agent] Answer done, content length:', message.content?.length || 0, 'answerEvent.content length:', answerEvent.content?.length || 0); // 完成 - 关闭所有状态 loading.value = false; @@ -735,9 +810,19 @@ const handleAgentChunk = (data) => { // 标题生成已改为异步事件推送,不再需要在这里手动调用 // 如果标题还未生成,前端会通过 SSE 事件接收 + } else if (data.done && answerEvent.done) { + console.log('[Answer] Ignoring duplicate done event, current content preserved:', answerEvent.content?.length || 0); } break; + case 'complete': + // 整个流式响应完成事件 - 确保状态正确关闭 + console.log('[Agent] Complete event received'); + loading.value = false; + isReplying.value = false; + // 不清空 message.content,保持已有内容 + break; + case 'stop': // 停止事件 - 添加到事件流并标记对话完成 console.log('[Agent] Stop event received'); diff --git a/internal/agent/engine.go b/internal/agent/engine.go index 46739685..3c923af1 100644 --- a/internal/agent/engine.go +++ b/internal/agent/engine.go @@ -688,12 +688,13 @@ func (e *AgentEngine) streamThinkingToEventBus( iteration int, sessionID string, ) (*types.ChatResponse, error) { - logger.Infof(ctx, "[Agent][Thinking][Iteration-%d] Starting thinking stream with temperature=%.2f, tools=%d", - iteration+1, e.config.Temperature, len(tools)) + logger.Infof(ctx, "[Agent][Thinking][Iteration-%d] Starting thinking stream with temperature=%.2f, tools=%d, thinking=%v", + iteration+1, e.config.Temperature, len(tools), e.config.Thinking) opts := &chat.ChatOptions{ Temperature: e.config.Temperature, Tools: tools, + Thinking: e.config.Thinking, } logger.Debug(context.Background(), "[Agent] streamLLM opts tool_choice=auto temperature=", e.config.Temperature) @@ -832,7 +833,7 @@ func (e *AgentEngine) streamFinalAnswerToEventBus( fullAnswer, _, err := e.streamLLMToEventBus( ctx, messages, - &chat.ChatOptions{Temperature: e.config.Temperature}, + &chat.ChatOptions{Temperature: e.config.Temperature, Thinking: e.config.Thinking}, func(chunk *types.StreamResponse, fullContent string) { if chunk.Content != "" { logger.Debugf(ctx, "[Agent][FinalAnswer] Emitting answer chunk: %d chars", len(chunk.Content)) diff --git a/internal/application/service/chat_pipline/chat_completion_stream.go b/internal/application/service/chat_pipline/chat_completion_stream.go index 507415e1..f618d6c9 100644 --- a/internal/application/service/chat_pipline/chat_completion_stream.go +++ b/internal/application/service/chat_pipline/chat_completion_stream.go @@ -99,9 +99,13 @@ func (p *PluginChatCompletionStream) OnEvent(ctx context.Context, }) // Start goroutine to consume channel and emit events directly + // For non-agent mode, thinking content is embedded in answer stream with tags + // This ensures consistent display between streaming and history loading go func() { answerID := fmt.Sprintf("%s-answer", uuid.New().String()[:8]) var finalContent string + var thinkingStarted bool + var thinkingEnded bool for response := range responseChan { // Handle error responses from the stream @@ -122,8 +126,53 @@ func (p *PluginChatCompletionStream) OnEvent(ctx context.Context, continue } + // For non-agent mode: embed thinking content with tags in answer stream + // This ensures the frontend uses deepThink.vue component consistently + if response.ResponseType == types.ResponseTypeThinking { + content := response.Content + // Add tag at the beginning of thinking content + if !thinkingStarted { + content = "" + content + thinkingStarted = true + } + // Add tag at the end of thinking content + if response.Done && !thinkingEnded { + content = content + "" + thinkingEnded = true + } + finalContent += content + if err := eventBus.Emit(ctx, types.Event{ + ID: answerID, + Type: types.EventType(event.EventAgentFinalAnswer), + SessionID: chatManage.SessionID, + Data: event.AgentFinalAnswerData{ + Content: content, + Done: false, // Thinking is not the final answer + }, + }); err != nil { + logger.Errorf(ctx, "Failed to emit thinking as answer event: %v", err) + } + continue + } + // Emit event for each answer chunk if response.ResponseType == types.ResponseTypeAnswer { + // If we had thinking but it wasn't explicitly ended, close the think tag + if thinkingStarted && !thinkingEnded { + thinkingEnded = true + finalContent += "" + if err := eventBus.Emit(ctx, types.Event{ + ID: answerID, + Type: types.EventType(event.EventAgentFinalAnswer), + SessionID: chatManage.SessionID, + Data: event.AgentFinalAnswerData{ + Content: "", + Done: false, + }, + }); err != nil { + logger.Errorf(ctx, "Failed to emit think close tag: %v", err) + } + } finalContent += response.Content if err := eventBus.Emit(ctx, types.Event{ ID: answerID, diff --git a/internal/application/service/session.go b/internal/application/service/session.go index d83136c4..fd6b8563 100644 --- a/internal/application/service/session.go +++ b/internal/application/service/session.go @@ -487,6 +487,12 @@ func (s *sessionService) KnowledgeQA( summaryConfig.MaxCompletionTokens = customAgent.Config.MaxCompletionTokens logger.Infof(ctx, "Using custom agent's max_completion_tokens: %d", customAgent.Config.MaxCompletionTokens) } + // Override thinking mode from agent config + // Agent-level thinking setting takes full control (no global fallback) + summaryConfig.Thinking = customAgent.Config.Thinking + if customAgent.Config.Thinking != nil { + logger.Infof(ctx, "Using custom agent's thinking: %v", *customAgent.Config.Thinking) + } // Override retrieval strategy settings if customAgent.Config.EmbeddingTopK > 0 { embeddingTopK = customAgent.Config.EmbeddingTopK @@ -1104,6 +1110,7 @@ func (s *sessionService) AgentQA( HistoryTurns: customAgent.Config.HistoryTurns, MCPSelectionMode: customAgent.Config.MCPSelectionMode, MCPServices: customAgent.Config.MCPServices, + Thinking: customAgent.Config.Thinking, } // Resolve knowledge bases: request-level @ mentions take priority over agent config diff --git a/internal/handler/session/helpers.go b/internal/handler/session/helpers.go index 49a6601f..d0cd6614 100644 --- a/internal/handler/session/helpers.go +++ b/internal/handler/session/helpers.go @@ -96,14 +96,14 @@ func buildStreamResponse(evt interfaces.StreamEvent, requestID string) *types.St } // sendCompletionEvent sends a final completion event to the client +// NOTE: This is now a no-op because: +// 1. The 'complete' event from handleComplete already signals stream completion +// 2. Sending an extra empty 'answer' event with done:true causes frontend issues +// (multiple done events can confuse state management) +// The frontend should use 'complete' response_type to detect stream completion func sendCompletionEvent(c *gin.Context, requestID string) { - c.SSEvent("message", &types.StreamResponse{ - ID: requestID, - ResponseType: types.ResponseTypeAnswer, - Content: "", - Done: true, - }) - c.Writer.Flush() + // Intentionally empty - completion is signaled by the 'complete' event + // which is already sent before this function is called } // createAgentQueryEvent creates a standard agent query event diff --git a/internal/handler/session/qa.go b/internal/handler/session/qa.go index cb2ff297..44360dd3 100644 --- a/internal/handler/session/qa.go +++ b/internal/handler/session/qa.go @@ -323,6 +323,10 @@ func (h *Handler) executeNormalModeQA(reqCtx *qaRequestContext, generateTitle bo streamCtx := h.setupSSEStream(reqCtx, generateTitle) // Setup completion handler for normal mode + // Note: Thinking content is now embedded in answer stream with tags + // by chat_completion_stream.go, so we don't need separate thinking event handling + var completionHandled bool // Prevent duplicate completion handling + streamCtx.eventBus.On(event.EventAgentFinalAnswer, func(ctx context.Context, evt event.Event) error { data, ok := evt.Data.(event.AgentFinalAnswerData) if !ok { @@ -330,14 +334,22 @@ func (h *Handler) executeNormalModeQA(reqCtx *qaRequestContext, generateTitle bo } streamCtx.assistantMessage.Content += data.Content if data.Done { + // Prevent duplicate completion handling + if completionHandled { + return nil + } + completionHandled = true + logger.Infof(streamCtx.asyncCtx, "Knowledge QA service completed for session: %s", sessionID) + // Content already contains ... tags from chat_completion_stream.go h.completeAssistantMessage(streamCtx.asyncCtx, streamCtx.assistantMessage) + // Emit EventAgentComplete - this will trigger handleComplete which sends the SSE complete event + // Note: Don't cancel context here, let the SSE handler close naturally after receiving the complete event streamCtx.eventBus.Emit(streamCtx.asyncCtx, event.Event{ Type: event.EventAgentComplete, SessionID: sessionID, Data: event.AgentCompleteData{FinalAnswer: streamCtx.assistantMessage.Content}, }) - streamCtx.cancel() } return nil }) diff --git a/internal/models/chat/chat.go b/internal/models/chat/chat.go index a0cb3489..781804fb 100644 --- a/internal/models/chat/chat.go +++ b/internal/models/chat/chat.go @@ -6,6 +6,7 @@ import ( "fmt" "strings" + "github.com/Tencent/WeKnora/internal/models/provider" "github.com/Tencent/WeKnora/internal/models/utils/ollama" "github.com/Tencent/WeKnora/internal/types" ) @@ -87,18 +88,41 @@ type ChatConfig struct { // NewChat 创建聊天实例 func NewChat(config *ChatConfig, ollamaService *ollama.OllamaService) (Chat, error) { - var chat Chat - var err error switch strings.ToLower(string(config.Source)) { case string(types.ModelSourceLocal): - chat, err = NewOllamaChat(config, ollamaService) - if err != nil { - return nil, err - } - return chat, nil + return NewOllamaChat(config, ollamaService) case string(types.ModelSourceRemote): - return NewRemoteAPIChat(config) + return NewRemoteChat(config) default: return nil, fmt.Errorf("unsupported chat model source: %s", config.Source) } } + +// NewRemoteChat 根据 provider 创建远程聊天实例 +func NewRemoteChat(config *ChatConfig) (Chat, error) { + providerName := provider.ProviderName(config.Provider) + if providerName == "" { + providerName = provider.DetectProvider(config.BaseURL) + } + + switch providerName { + case provider.ProviderLKEAP: + // LKEAP 有特殊的 thinking 参数格式 + return NewLKEAPChat(config) + case provider.ProviderAliyun: + // 检查是否为 Qwen3 模型(需要特殊处理 enable_thinking) + if provider.IsQwen3Model(config.ModelName) { + return NewQwenChat(config) + } + return NewRemoteAPIChat(config) + case provider.ProviderDeepSeek: + // DeepSeek 不支持 tool_choice + return NewDeepSeekChat(config) + case provider.ProviderGeneric: + // Generic provider (如 vLLM) 使用 ChatTemplateKwargs + return NewGenericChat(config) + default: + // 其他 provider 使用标准 OpenAI 兼容实现 + return NewRemoteAPIChat(config) + } +} diff --git a/internal/models/chat/lkeap.go b/internal/models/chat/lkeap.go new file mode 100644 index 00000000..395a7e60 --- /dev/null +++ b/internal/models/chat/lkeap.go @@ -0,0 +1,77 @@ +package chat + +import ( + "strings" + + "github.com/Tencent/WeKnora/internal/models/provider" + "github.com/sashabaranov/go-openai" +) + +// LKEAPChat 腾讯云知识引擎原子能力 (LKEAP) 聊天实现 +// 支持 DeepSeek-R1, DeepSeek-V3 系列模型,具备思维链能力 +// 参考:https://cloud.tencent.com/document/product/1772/115963 +// +// 与标准 OpenAI API 的区别: +// 1. thinking 参数格式不同:LKEAP 使用 {"type": "enabled"/"disabled"} +// 2. 仅 DeepSeek V3.x 系列需要显式设置 thinking 参数,R1 系列默认开启 +type LKEAPChat struct { + *RemoteAPIChat +} + +// LKEAPThinkingConfig 思维链配置(LKEAP 特有格式) +type LKEAPThinkingConfig struct { + Type string `json:"type"` // "enabled" 或 "disabled" +} + +// LKEAPChatCompletionRequest LKEAP 自定义请求结构体 +type LKEAPChatCompletionRequest struct { + openai.ChatCompletionRequest + Thinking *LKEAPThinkingConfig `json:"thinking,omitempty"` // 思维链开关(仅 V3.x 系列) +} + +// NewLKEAPChat 创建 LKEAP 聊天实例 +func NewLKEAPChat(config *ChatConfig) (*LKEAPChat, error) { + // 确保 provider 设置正确 + config.Provider = string(provider.ProviderLKEAP) + + remoteChat, err := NewRemoteAPIChat(config) + if err != nil { + return nil, err + } + + chat := &LKEAPChat{ + RemoteAPIChat: remoteChat, + } + + // 设置请求自定义器,添加 LKEAP 特有的 thinking 参数 + remoteChat.SetRequestCustomizer(chat.customizeRequest) + + return chat, nil +} + +// isDeepSeekV3Model 检查是否为 DeepSeek V3.x 系列模型 +func (c *LKEAPChat) isDeepSeekV3Model() bool { + return strings.Contains(strings.ToLower(c.GetModelName()), "deepseek-v3") +} + +// customizeRequest 自定义 LKEAP 请求 +func (c *LKEAPChat) customizeRequest(req *openai.ChatCompletionRequest, opts *ChatOptions, isStream bool) (any, bool) { + // 仅对 DeepSeek V3.x 系列模型需要特殊处理 thinking 参数 + // R1 系列模型默认开启思维链,无需额外参数 + if !c.isDeepSeekV3Model() || opts == nil || opts.Thinking == nil { + return nil, false // 使用标准请求 + } + + // 构建 LKEAP 特有请求 + lkeapReq := LKEAPChatCompletionRequest{ + ChatCompletionRequest: *req, + } + + thinkingType := "disabled" + if *opts.Thinking { + thinkingType = "enabled" + } + lkeapReq.Thinking = &LKEAPThinkingConfig{Type: thinkingType} + + return lkeapReq, true // 使用原始 HTTP 请求 +} diff --git a/internal/models/chat/provider_chat.go b/internal/models/chat/provider_chat.go new file mode 100644 index 00000000..71837dba --- /dev/null +++ b/internal/models/chat/provider_chat.go @@ -0,0 +1,82 @@ +package chat + +import ( + "context" + + "github.com/Tencent/WeKnora/internal/logger" + "github.com/Tencent/WeKnora/internal/models/provider" + "github.com/sashabaranov/go-openai" +) + +// DeepSeekChat DeepSeek 模型聊天实现 +// DeepSeek 模型不支持 tool_choice 参数 +type DeepSeekChat struct { + *RemoteAPIChat +} + +// NewDeepSeekChat 创建 DeepSeek 聊天实例 +func NewDeepSeekChat(config *ChatConfig) (*DeepSeekChat, error) { + config.Provider = string(provider.ProviderDeepSeek) + + remoteChat, err := NewRemoteAPIChat(config) + if err != nil { + return nil, err + } + + chat := &DeepSeekChat{ + RemoteAPIChat: remoteChat, + } + + // 设置请求自定义器 + remoteChat.SetRequestCustomizer(chat.customizeRequest) + + return chat, nil +} + +// customizeRequest 自定义 DeepSeek 请求 +func (c *DeepSeekChat) customizeRequest(req *openai.ChatCompletionRequest, opts *ChatOptions, isStream bool) (any, bool) { + // DeepSeek 模型不支持 tool_choice,需要清除 + if opts != nil && opts.ToolChoice != "" { + logger.Infof(context.Background(), "deepseek model, skip tool_choice") + req.ToolChoice = nil + } + return nil, false +} + +// GenericChat 通用 OpenAI 兼容实现(如 vLLM) +// 支持 ChatTemplateKwargs 参数 +type GenericChat struct { + *RemoteAPIChat +} + +// NewGenericChat 创建通用聊天实例 +func NewGenericChat(config *ChatConfig) (*GenericChat, error) { + config.Provider = string(provider.ProviderGeneric) + + remoteChat, err := NewRemoteAPIChat(config) + if err != nil { + return nil, err + } + + chat := &GenericChat{ + RemoteAPIChat: remoteChat, + } + + // 设置请求自定义器 + remoteChat.SetRequestCustomizer(chat.customizeRequest) + + return chat, nil +} + +// customizeRequest 自定义 Generic 请求 +func (c *GenericChat) customizeRequest(req *openai.ChatCompletionRequest, opts *ChatOptions, isStream bool) (any, bool) { + // Generic provider(如 vLLM)使用 ChatTemplateKwargs 传递 thinking 参数 + thinking := false + if opts != nil && opts.Thinking != nil { + thinking = *opts.Thinking + } + req.ChatTemplateKwargs = map[string]interface{}{ + "enable_thinking": thinking, + } + return nil, false // 使用标准请求(已修改) +} diff --git a/internal/models/chat/qwen.go b/internal/models/chat/qwen.go new file mode 100644 index 00000000..9fca5d09 --- /dev/null +++ b/internal/models/chat/qwen.go @@ -0,0 +1,62 @@ +package chat + +import ( + "github.com/Tencent/WeKnora/internal/models/provider" + "github.com/sashabaranov/go-openai" +) + +// QwenChat 阿里云 Qwen 模型聊天实现 +// Qwen3 模型需要特殊处理 enable_thinking 参数 +type QwenChat struct { + *RemoteAPIChat +} + +// QwenChatCompletionRequest Qwen 模型的自定义请求结构体 +type QwenChatCompletionRequest struct { + openai.ChatCompletionRequest + EnableThinking *bool `json:"enable_thinking,omitempty"` +} + +// NewQwenChat 创建 Qwen 聊天实例 +func NewQwenChat(config *ChatConfig) (*QwenChat, error) { + config.Provider = string(provider.ProviderAliyun) + + remoteChat, err := NewRemoteAPIChat(config) + if err != nil { + return nil, err + } + + chat := &QwenChat{ + RemoteAPIChat: remoteChat, + } + + // 设置请求自定义器 + remoteChat.SetRequestCustomizer(chat.customizeRequest) + + return chat, nil +} + +// isQwen3Model 检查是否为 Qwen3 模型 +func (c *QwenChat) isQwen3Model() bool { + return provider.IsQwen3Model(c.GetModelName()) +} + +// customizeRequest 自定义 Qwen 请求 +func (c *QwenChat) customizeRequest(req *openai.ChatCompletionRequest, opts *ChatOptions, isStream bool) (any, bool) { + // 仅 Qwen3 模型需要特殊处理 + if !c.isQwen3Model() { + return nil, false + } + + // 非流式请求需要显式禁用 thinking + if !isStream { + qwenReq := QwenChatCompletionRequest{ + ChatCompletionRequest: *req, + } + enableThinking := false + qwenReq.EnableThinking = &enableThinking + return qwenReq, true + } + + return nil, false +} diff --git a/internal/models/chat/remote_api.go b/internal/models/chat/remote_api.go index 4193f7b1..0a5fec0d 100644 --- a/internal/models/chat/remote_api.go +++ b/internal/models/chat/remote_api.go @@ -5,6 +5,7 @@ import ( "context" "encoding/json" "fmt" + "io" "net/http" "github.com/Tencent/WeKnora/internal/logger" @@ -13,23 +14,22 @@ import ( "github.com/sashabaranov/go-openai" ) -// RemoteAPIChat 实现了基于的聊天 +// RemoteAPIChat 实现了基于 OpenAI 兼容 API 的聊天 +// 这是一个通用实现,不包含任何 provider 特定的逻辑 type RemoteAPIChat struct { modelName string client *openai.Client modelID string baseURL string apiKey string - provider provider.ProviderName // Provider identifier for routing + provider provider.ProviderName + + // requestCustomizer 允许子类自定义请求 + // 返回自定义请求体(如果为 nil 则使用标准请求)和是否需要使用原始 HTTP 请求 + requestCustomizer func(req *openai.ChatCompletionRequest, opts *ChatOptions, isStream bool) (customReq any, useRawHTTP bool) } -// QwenChatCompletionRequest 用于 qwen 模型的自定义请求结构体 -type QwenChatCompletionRequest struct { - openai.ChatCompletionRequest - EnableThinking *bool `json:"enable_thinking,omitempty"` // qwen 模型专用字段 -} - -// NewRemoteAPIChat 调用远程API 聊天实例 +// NewRemoteAPIChat 创建远程 API 聊天实例 func NewRemoteAPIChat(chatConfig *ChatConfig) (*RemoteAPIChat, error) { apiKey := chatConfig.APIKey config := openai.DefaultConfig(apiKey) @@ -37,7 +37,6 @@ func NewRemoteAPIChat(chatConfig *ChatConfig) (*RemoteAPIChat, error) { config.BaseURL = baseURL } - // Detect or use configured provider providerName := provider.ProviderName(chatConfig.Provider) if providerName == "" { providerName = provider.DetectProvider(chatConfig.BaseURL) @@ -53,20 +52,23 @@ func NewRemoteAPIChat(chatConfig *ChatConfig) (*RemoteAPIChat, error) { }, nil } -// convertMessages 转换消息格式为OpenAI格式 -func (c *RemoteAPIChat) convertMessages(messages []Message) []openai.ChatCompletionMessage { +// SetRequestCustomizer 设置请求自定义器 +func (c *RemoteAPIChat) SetRequestCustomizer(customizer func(req *openai.ChatCompletionRequest, opts *ChatOptions, isStream bool) (any, bool)) { + c.requestCustomizer = customizer +} + +// ConvertMessages 转换消息格式为 OpenAI 格式(导出供子类使用) +func (c *RemoteAPIChat) ConvertMessages(messages []Message) []openai.ChatCompletionMessage { openaiMessages := make([]openai.ChatCompletionMessage, 0, len(messages)) for _, msg := range messages { openaiMsg := openai.ChatCompletionMessage{ Role: msg.Role, } - // 处理内容:对于 assistant 角色,内容可能为空(当有 tool_calls 时) if msg.Content != "" { openaiMsg.Content = msg.Content } - // 处理 tool calls(assistant 角色) if len(msg.ToolCalls) > 0 { openaiMsg.ToolCalls = make([]openai.ToolCall, 0, len(msg.ToolCalls)) for _, tc := range msg.ToolCalls { @@ -82,7 +84,6 @@ func (c *RemoteAPIChat) convertMessages(messages []Message) []openai.ChatComplet } } - // 处理 tool 角色消息(工具返回结果) if msg.Role == "tool" { openaiMsg.ToolCallID = msg.ToolCallID openaiMsg.Name = msg.Name @@ -93,44 +94,14 @@ func (c *RemoteAPIChat) convertMessages(messages []Message) []openai.ChatComplet return openaiMessages } -// isQwenModel 检查是否为 qwen 模型 -func (c *RemoteAPIChat) isAliyunQwen3Model() bool { - return c.provider == provider.ProviderAliyun && provider.IsQwen3Model(c.modelName) -} - -// isDeepSeekModel 检查是否为 DeepSeek 模型 -func (c *RemoteAPIChat) isDeepSeekModel() bool { - return provider.IsDeepSeekModel(c.modelName) -} - -// buildQwenChatCompletionRequest 构建 qwen 模型的聊天请求参数 -func (c *RemoteAPIChat) buildQwenChatCompletionRequest(messages []Message, - opts *ChatOptions, isStream bool, -) QwenChatCompletionRequest { - req := QwenChatCompletionRequest{ - ChatCompletionRequest: c.buildChatCompletionRequest(messages, opts, isStream), - } - - // 对于 qwen 模型,在非流式调用中强制设置 enable_thinking: false - if !isStream { - enableThinking := false - req.EnableThinking = &enableThinking - } - return req -} - -// buildChatCompletionRequest 构建聊天请求参数 -func (c *RemoteAPIChat) buildChatCompletionRequest(messages []Message, - opts *ChatOptions, isStream bool, -) openai.ChatCompletionRequest { +// BuildChatCompletionRequest 构建标准聊天请求参数(导出供子类使用) +func (c *RemoteAPIChat) BuildChatCompletionRequest(messages []Message, opts *ChatOptions, isStream bool) openai.ChatCompletionRequest { req := openai.ChatCompletionRequest{ Model: c.modelName, - Messages: c.convertMessages(messages), + Messages: c.ConvertMessages(messages), Stream: isStream, } - thinking := false - // 添加可选参数 if opts != nil { if opts.Temperature > 0 { req.Temperature = float32(opts.Temperature) @@ -150,11 +121,8 @@ func (c *RemoteAPIChat) buildChatCompletionRequest(messages []Message, if opts.PresencePenalty > 0 { req.PresencePenalty = float32(opts.PresencePenalty) } - if opts.Thinking != nil { - thinking = *opts.Thinking - } - // 处理 Tools(函数定义) + // 处理 Tools if len(opts.Tools) > 0 { req.Tools = make([]openai.Tool, 0, len(opts.Tools)) for _, tool := range opts.Tools { @@ -166,39 +134,24 @@ func (c *RemoteAPIChat) buildChatCompletionRequest(messages []Message, Description: tool.Function.Description, }, } - // 转换 Parameters (map[string]interface{} -> JSON Schema) if tool.Function.Parameters != nil { - // Parameters 已经是 JSON Schema 格式的 map,直接使用 openaiTool.Function.Parameters = tool.Function.Parameters } req.Tools = append(req.Tools, openaiTool) } } - // 处理 ToolChoice - // ToolChoice 可以是字符串或 ToolChoice 对象 - // 对于 "auto", "none", "required" 直接使用字符串 - // 对于特定工具名称,使用 ToolChoice 对象 - // 注意:某些模型(如 DeepSeek)不支持 tool_choice,需要跳过设置 + // 处理 ToolChoice(标准实现) if opts.ToolChoice != "" { - // DeepSeek 模型不支持 tool_choice,跳过设置(默认行为会自动使用工具) - if c.isDeepSeekModel() { - // 对于 DeepSeek,不设置 tool_choice,让 API 使用默认行为 - // 如果有 tools,DeepSeek 会自动使用 - logger.Infof(context.Background(), "deepseek model, skip tool_choice") - } else { - switch opts.ToolChoice { - case "none", "required", "auto": - // 直接使用字符串 - req.ToolChoice = opts.ToolChoice - default: - // 特定工具名称,使用 ToolChoice 对象 - req.ToolChoice = openai.ToolChoice{ - Type: "function", - Function: openai.ToolFunction{ - Name: opts.ToolChoice, - }, - } + switch opts.ToolChoice { + case "none", "required", "auto": + req.ToolChoice = opts.ToolChoice + default: + req.ToolChoice = openai.ToolChoice{ + Type: "function", + Function: openai.ToolFunction{ + Name: opts.ToolChoice, + }, } } } @@ -211,50 +164,80 @@ func (c *RemoteAPIChat) buildChatCompletionRequest(messages []Message, } } - // ChatTemplateKwargs is only supported by custom backends like vLLM. - // Official APIs (OpenAI, Aliyun, Zhipu, etc.) do not support this parameter - // and will return 400 Bad Request if it's included. - if c.provider == provider.ProviderGeneric { - req.ChatTemplateKwargs = map[string]interface{}{ - "enable_thinking": thinking, - } - } - - // Log LLM request for debugging - if jsonData, err := json.MarshalIndent(req, "", " "); err == nil { - logger.Infof(context.Background(), "[LLM Request] model=%s, stream=%v, request:\n%s", c.modelName, isStream, string(jsonData)) - } - - // Log tools/functions separately for clarity - if len(req.Tools) > 0 { - toolNames := make([]string, 0, len(req.Tools)) - for _, tool := range req.Tools { - toolNames = append(toolNames, tool.Function.Name) - } - logger.Infof(context.Background(), "[LLM Request] tools_count=%d, tool_names=%v", len(req.Tools), toolNames) - } - return req } +// logRequest 记录请求日志 +func (c *RemoteAPIChat) logRequest(ctx context.Context, req any, isStream bool) { + if jsonData, err := json.MarshalIndent(req, "", " "); err == nil { + logger.Infof(ctx, "[LLM Request] model=%s, stream=%v, request:\n%s", c.modelName, isStream, string(jsonData)) + } +} + // Chat 进行非流式聊天 func (c *RemoteAPIChat) Chat(ctx context.Context, messages []Message, opts *ChatOptions) (*types.ChatResponse, error) { - // 如果是 qwen 模型,使用自定义请求 - if c.isAliyunQwen3Model() { - return c.chatWithQwen(ctx, messages, opts) + req := c.BuildChatCompletionRequest(messages, opts, false) + + // 检查是否需要自定义请求 + if c.requestCustomizer != nil { + customReq, useRawHTTP := c.requestCustomizer(&req, opts, false) + if useRawHTTP && customReq != nil { + return c.chatWithRawHTTP(ctx, customReq) + } } - // 构建请求参数 - req := c.buildChatCompletionRequest(messages, opts, false) + c.logRequest(ctx, req, false) - // 发送请求 resp, err := c.client.CreateChatCompletion(ctx, req) if err != nil { return nil, fmt.Errorf("create chat completion: %w", err) } + return c.parseCompletionResponse(&resp) +} + +// chatWithRawHTTP 使用原始 HTTP 请求进行聊天(供自定义请求使用) +func (c *RemoteAPIChat) chatWithRawHTTP(ctx context.Context, customReq any) (*types.ChatResponse, error) { + jsonData, err := json.Marshal(customReq) + if err != nil { + return nil, fmt.Errorf("marshal request: %w", err) + } + + logger.Infof(ctx, "[LLM Request] model=%s, raw HTTP request:\n%s", c.modelName, string(jsonData)) + + endpoint := c.baseURL + "/chat/completions" + httpReq, err := http.NewRequestWithContext(ctx, "POST", endpoint, bytes.NewBuffer(jsonData)) + if err != nil { + return nil, fmt.Errorf("create request: %w", err) + } + + httpReq.Header.Set("Content-Type", "application/json") + httpReq.Header.Set("Authorization", "Bearer "+c.apiKey) + + client := &http.Client{} + resp, err := client.Do(httpReq) + if err != nil { + return nil, fmt.Errorf("send request: %w", err) + } + defer resp.Body.Close() + + if resp.StatusCode != http.StatusOK { + body, _ := io.ReadAll(resp.Body) + return nil, fmt.Errorf("API request failed with status %d: %s", resp.StatusCode, string(body)) + } + + var chatResp openai.ChatCompletionResponse + if err := json.NewDecoder(resp.Body).Decode(&chatResp); err != nil { + return nil, fmt.Errorf("decode response: %w", err) + } + + return c.parseCompletionResponse(&chatResp) +} + +// parseCompletionResponse 解析非流式响应 +func (c *RemoteAPIChat) parseCompletionResponse(resp *openai.ChatCompletionResponse) (*types.ChatResponse, error) { if len(resp.Choices) == 0 { - return nil, fmt.Errorf("no response from OpenAI") + return nil, fmt.Errorf("no response from API") } choice := resp.Choices[0] @@ -272,91 +255,6 @@ func (c *RemoteAPIChat) Chat(ctx context.Context, messages []Message, opts *Chat }, } - // 转换 Tool Calls - if len(choice.Message.ToolCalls) > 0 { - response.ToolCalls = make([]types.LLMToolCall, 0, len(choice.Message.ToolCalls)) - for _, tc := range choice.Message.ToolCalls { - response.ToolCalls = append(response.ToolCalls, types.LLMToolCall{ - ID: tc.ID, - Type: string(tc.Type), - Function: types.FunctionCall{ - Name: tc.Function.Name, - Arguments: tc.Function.Arguments, - }, - }) - } - } - - return response, nil -} - -// chatWithQwen 使用自定义请求处理 qwen 模型 -func (c *RemoteAPIChat) chatWithQwen( - ctx context.Context, - messages []Message, - opts *ChatOptions, -) (*types.ChatResponse, error) { - // 构建 qwen 请求参数 - req := c.buildQwenChatCompletionRequest(messages, opts, false) - - // 序列化请求 - jsonData, err := json.Marshal(req) - if err != nil { - return nil, fmt.Errorf("marshal request: %w", err) - } - - // 构建 URL - endpoint := c.baseURL + "/chat/completions" - - // 创建 HTTP 请求 - httpReq, err := http.NewRequestWithContext(ctx, "POST", endpoint, bytes.NewBuffer(jsonData)) - if err != nil { - return nil, fmt.Errorf("create request: %w", err) - } - - // 设置请求头 - httpReq.Header.Set("Content-Type", "application/json") - httpReq.Header.Set("Authorization", "Bearer "+c.apiKey) - - // 发送请求 - client := &http.Client{} - resp, err := client.Do(httpReq) - if err != nil { - return nil, fmt.Errorf("send request: %w", err) - } - defer resp.Body.Close() - - // 检查响应状态 - if resp.StatusCode != http.StatusOK { - return nil, fmt.Errorf("API request failed with status: %d", resp.StatusCode) - } - - // 解析响应 - var chatResp openai.ChatCompletionResponse - if err := json.NewDecoder(resp.Body).Decode(&chatResp); err != nil { - return nil, fmt.Errorf("decode response: %w", err) - } - - if len(chatResp.Choices) == 0 { - return nil, fmt.Errorf("no response from API") - } - - choice := chatResp.Choices[0] - response := &types.ChatResponse{ - Content: choice.Message.Content, - FinishReason: string(choice.FinishReason), - Usage: struct { - PromptTokens int `json:"prompt_tokens"` - CompletionTokens int `json:"completion_tokens"` - TotalTokens int `json:"total_tokens"` - }{ - PromptTokens: chatResp.Usage.PromptTokens, - CompletionTokens: chatResp.Usage.CompletionTokens, - TotalTokens: chatResp.Usage.TotalTokens, - }, - } - - // 转换 Tool Calls if len(choice.Message.ToolCalls) > 0 { response.ToolCalls = make([]types.LLMToolCall, 0, len(choice.Message.ToolCalls)) for _, tc := range choice.Message.ToolCalls { @@ -375,162 +273,295 @@ func (c *RemoteAPIChat) chatWithQwen( } // ChatStream 进行流式聊天 -func (c *RemoteAPIChat) ChatStream(ctx context.Context, - messages []Message, opts *ChatOptions, -) (<-chan types.StreamResponse, error) { - // 构建请求参数 - req := c.buildChatCompletionRequest(messages, opts, true) +func (c *RemoteAPIChat) ChatStream(ctx context.Context, messages []Message, opts *ChatOptions) (<-chan types.StreamResponse, error) { + req := c.BuildChatCompletionRequest(messages, opts, true) + + // 检查是否需要自定义请求 + if c.requestCustomizer != nil { + customReq, useRawHTTP := c.requestCustomizer(&req, opts, true) + if useRawHTTP && customReq != nil { + return c.chatStreamWithRawHTTP(ctx, customReq) + } + } + + c.logRequest(ctx, req, true) - // 创建流式响应通道 streamChan := make(chan types.StreamResponse) - // 启动流式请求 stream, err := c.client.CreateChatCompletionStream(ctx, req) if err != nil { close(streamChan) return nil, fmt.Errorf("create chat completion stream: %w", err) } - // 在后台处理流式响应 - go func() { - defer close(streamChan) - defer stream.Close() - - toolCallMap := make(map[int]*types.LLMToolCall) - lastFunctionName := make(map[int]string) - nameNotified := make(map[int]bool) - - buildOrderedToolCalls := func() []types.LLMToolCall { - if len(toolCallMap) == 0 { - return nil - } - result := make([]types.LLMToolCall, 0, len(toolCallMap)) - for i := 0; i < len(toolCallMap); i++ { - if tc, ok := toolCallMap[i]; ok && tc != nil { - result = append(result, *tc) - } - } - if len(result) == 0 { - return nil - } - return result - } - - for { - response, err := stream.Recv() - if err != nil { - // Check if it's a normal end of stream (io.EOF) - if err.Error() == "EOF" { - // Normal end of stream, send final response with collected tool calls - streamChan <- types.StreamResponse{ - ResponseType: types.ResponseTypeAnswer, - Content: "", - Done: true, - ToolCalls: buildOrderedToolCalls(), - } - } else { - // Actual error, send error response - streamChan <- types.StreamResponse{ - ResponseType: types.ResponseTypeError, - Content: err.Error(), - Done: true, - } - } - return - } - - if len(response.Choices) > 0 { - delta := response.Choices[0].Delta - isDone := string(response.Choices[0].FinishReason) != "" - - // 收集 tool calls(流式响应中 tool calls 可能分多次返回) - if len(delta.ToolCalls) > 0 { - for _, tc := range delta.ToolCalls { - // 检查是否已经存在该 tool call(通过 index) - var toolCallIndex int - if tc.Index != nil { - toolCallIndex = *tc.Index - } - toolCallEntry, exists := toolCallMap[toolCallIndex] - if !exists || toolCallEntry == nil { - toolCallEntry = &types.LLMToolCall{ - Type: string(tc.Type), - Function: types.FunctionCall{ - Name: "", - Arguments: "", - }, - } - toolCallMap[toolCallIndex] = toolCallEntry - } - - // 更新 ID、类型 - if tc.ID != "" { - toolCallEntry.ID = tc.ID - } - if tc.Type != "" { - toolCallEntry.Type = string(tc.Type) - } - - // 累积函数名称(可能分多次返回) - if tc.Function.Name != "" { - toolCallEntry.Function.Name += tc.Function.Name - } - - // 累积参数(可能为部分 JSON) - argsUpdated := false - if tc.Function.Arguments != "" { - toolCallEntry.Function.Arguments += tc.Function.Arguments - argsUpdated = true - } - - currName := toolCallEntry.Function.Name - if currName != "" && - currName == lastFunctionName[toolCallIndex] && - argsUpdated && - !nameNotified[toolCallIndex] && - toolCallEntry.ID != "" { - streamChan <- types.StreamResponse{ - ResponseType: types.ResponseTypeToolCall, - Content: "", - Done: false, - Data: map[string]interface{}{ - "tool_name": currName, - "tool_call_id": toolCallEntry.ID, - }, - } - nameNotified[toolCallIndex] = true - } - - lastFunctionName[toolCallIndex] = currName - } - } - - // 发送内容块 - if delta.Content != "" { - streamChan <- types.StreamResponse{ - ResponseType: types.ResponseTypeAnswer, - Content: delta.Content, - Done: isDone, - ToolCalls: buildOrderedToolCalls(), - } - } - - // 如果是最后一次响应,确保发送包含所有 tool calls 的响应 - if isDone && len(toolCallMap) > 0 { - streamChan <- types.StreamResponse{ - ResponseType: types.ResponseTypeAnswer, - Content: "", - Done: true, - ToolCalls: buildOrderedToolCalls(), - } - } - } - } - }() + go c.processStream(ctx, stream, streamChan) return streamChan, nil } +// chatStreamWithRawHTTP 使用原始 HTTP 请求进行流式聊天 +func (c *RemoteAPIChat) chatStreamWithRawHTTP(ctx context.Context, customReq any) (<-chan types.StreamResponse, error) { + jsonData, err := json.Marshal(customReq) + if err != nil { + return nil, fmt.Errorf("marshal request: %w", err) + } + + logger.Infof(ctx, "[LLM Stream] model=%s, raw HTTP request:\n%s", c.modelName, string(jsonData)) + + endpoint := c.baseURL + "/chat/completions" + httpReq, err := http.NewRequestWithContext(ctx, "POST", endpoint, bytes.NewBuffer(jsonData)) + if err != nil { + return nil, fmt.Errorf("create request: %w", err) + } + + httpReq.Header.Set("Content-Type", "application/json") + httpReq.Header.Set("Authorization", "Bearer "+c.apiKey) + httpReq.Header.Set("Accept", "text/event-stream") + + client := &http.Client{} + resp, err := client.Do(httpReq) + if err != nil { + return nil, fmt.Errorf("send request: %w", err) + } + + if resp.StatusCode != http.StatusOK { + body, _ := io.ReadAll(resp.Body) + resp.Body.Close() + return nil, fmt.Errorf("API request failed with status %d: %s", resp.StatusCode, string(body)) + } + + streamChan := make(chan types.StreamResponse) + + go c.processRawHTTPStream(ctx, resp, streamChan) + + return streamChan, nil +} + +// processStream 处理 OpenAI SDK 流式响应 +func (c *RemoteAPIChat) processStream(ctx context.Context, stream *openai.ChatCompletionStream, streamChan chan types.StreamResponse) { + defer close(streamChan) + defer stream.Close() + + state := newStreamState() + + for { + response, err := stream.Recv() + if err != nil { + if err.Error() == "EOF" { + streamChan <- types.StreamResponse{ + ResponseType: types.ResponseTypeAnswer, + Content: "", + Done: true, + ToolCalls: state.buildOrderedToolCalls(), + } + } else { + streamChan <- types.StreamResponse{ + ResponseType: types.ResponseTypeError, + Content: err.Error(), + Done: true, + } + } + return + } + + if len(response.Choices) > 0 { + c.processStreamDelta(ctx, &response.Choices[0], state, streamChan) + } + } +} + +// processRawHTTPStream 处理原始 HTTP 流式响应 +func (c *RemoteAPIChat) processRawHTTPStream(ctx context.Context, resp *http.Response, streamChan chan types.StreamResponse) { + defer close(streamChan) + defer resp.Body.Close() + + state := newStreamState() + reader := NewSSEReader(resp.Body) + + for { + event, err := reader.ReadEvent() + if err != nil { + if err.Error() != "EOF" { + logger.Errorf(ctx, "Stream read error: %v", err) + streamChan <- types.StreamResponse{ + ResponseType: types.ResponseTypeError, + Content: err.Error(), + Done: true, + } + } + return + } + + if event == nil { + continue + } + + if event.Done { + streamChan <- types.StreamResponse{ + ResponseType: types.ResponseTypeAnswer, + Content: "", + Done: true, + ToolCalls: state.buildOrderedToolCalls(), + } + return + } + + if event.Data == nil { + continue + } + + var streamResp openai.ChatCompletionStreamResponse + if err := json.Unmarshal(event.Data, &streamResp); err != nil { + logger.Errorf(ctx, "Failed to parse stream response: %v", err) + continue + } + + if len(streamResp.Choices) > 0 { + c.processStreamDelta(ctx, &streamResp.Choices[0], state, streamChan) + } + } +} + +// streamState 流式处理状态 +type streamState struct { + toolCallMap map[int]*types.LLMToolCall + lastFunctionName map[int]string + nameNotified map[int]bool + hasThinking bool +} + +func newStreamState() *streamState { + return &streamState{ + toolCallMap: make(map[int]*types.LLMToolCall), + lastFunctionName: make(map[int]string), + nameNotified: make(map[int]bool), + hasThinking: false, + } +} + +func (s *streamState) buildOrderedToolCalls() []types.LLMToolCall { + if len(s.toolCallMap) == 0 { + return nil + } + result := make([]types.LLMToolCall, 0, len(s.toolCallMap)) + for i := 0; i < len(s.toolCallMap); i++ { + if tc, ok := s.toolCallMap[i]; ok && tc != nil { + result = append(result, *tc) + } + } + if len(result) == 0 { + return nil + } + return result +} + +// processStreamDelta 处理流式响应的单个 delta +func (c *RemoteAPIChat) processStreamDelta(ctx context.Context, choice *openai.ChatCompletionStreamChoice, state *streamState, streamChan chan types.StreamResponse) { + delta := choice.Delta + isDone := string(choice.FinishReason) != "" + + // 处理 tool calls + if len(delta.ToolCalls) > 0 { + c.processToolCallsDelta(delta.ToolCalls, state, streamChan) + } + + // 发送思考内容(ReasoningContent,支持 DeepSeek 等模型) + if delta.ReasoningContent != "" { + state.hasThinking = true + streamChan <- types.StreamResponse{ + ResponseType: types.ResponseTypeThinking, + Content: delta.ReasoningContent, + Done: false, + } + } + + // 发送回答内容 + if delta.Content != "" { + // If we had thinking content and this is the first answer chunk, + // send a thinking done event first + if state.hasThinking { + streamChan <- types.StreamResponse{ + ResponseType: types.ResponseTypeThinking, + Content: "", + Done: true, + } + state.hasThinking = false // Only send once + } + streamChan <- types.StreamResponse{ + ResponseType: types.ResponseTypeAnswer, + Content: delta.Content, + Done: isDone, + ToolCalls: state.buildOrderedToolCalls(), + } + } + + if isDone && len(state.toolCallMap) > 0 { + streamChan <- types.StreamResponse{ + ResponseType: types.ResponseTypeAnswer, + Content: "", + Done: true, + ToolCalls: state.buildOrderedToolCalls(), + } + } +} + +// processToolCallsDelta 处理 tool calls 的增量更新 +func (c *RemoteAPIChat) processToolCallsDelta(toolCalls []openai.ToolCall, state *streamState, streamChan chan types.StreamResponse) { + for _, tc := range toolCalls { + var toolCallIndex int + if tc.Index != nil { + toolCallIndex = *tc.Index + } + toolCallEntry, exists := state.toolCallMap[toolCallIndex] + if !exists || toolCallEntry == nil { + toolCallEntry = &types.LLMToolCall{ + Type: string(tc.Type), + Function: types.FunctionCall{ + Name: "", + Arguments: "", + }, + } + state.toolCallMap[toolCallIndex] = toolCallEntry + } + + if tc.ID != "" { + toolCallEntry.ID = tc.ID + } + if tc.Type != "" { + toolCallEntry.Type = string(tc.Type) + } + if tc.Function.Name != "" { + toolCallEntry.Function.Name += tc.Function.Name + } + + argsUpdated := false + if tc.Function.Arguments != "" { + toolCallEntry.Function.Arguments += tc.Function.Arguments + argsUpdated = true + } + + currName := toolCallEntry.Function.Name + if currName != "" && + currName == state.lastFunctionName[toolCallIndex] && + argsUpdated && + !state.nameNotified[toolCallIndex] && + toolCallEntry.ID != "" { + streamChan <- types.StreamResponse{ + ResponseType: types.ResponseTypeToolCall, + Content: "", + Done: false, + Data: map[string]interface{}{ + "tool_name": currName, + "tool_call_id": toolCallEntry.ID, + }, + } + state.nameNotified[toolCallIndex] = true + } + + state.lastFunctionName[toolCallIndex] = currName + } +} + // GetModelName 获取模型名称 func (c *RemoteAPIChat) GetModelName() string { return c.modelName @@ -540,3 +571,18 @@ func (c *RemoteAPIChat) GetModelName() string { func (c *RemoteAPIChat) GetModelID() string { return c.modelID } + +// GetProvider 获取 provider 名称 +func (c *RemoteAPIChat) GetProvider() provider.ProviderName { + return c.provider +} + +// GetBaseURL 获取 baseURL +func (c *RemoteAPIChat) GetBaseURL() string { + return c.baseURL +} + +// GetAPIKey 获取 apiKey +func (c *RemoteAPIChat) GetAPIKey() string { + return c.apiKey +} diff --git a/internal/models/chat/sse_reader.go b/internal/models/chat/sse_reader.go new file mode 100644 index 00000000..45c6cf2b --- /dev/null +++ b/internal/models/chat/sse_reader.go @@ -0,0 +1,59 @@ +package chat + +import ( + "bufio" + "errors" + "io" + "strings" +) + +// SSEEvent 表示一个 Server-Sent Events 事件 +type SSEEvent struct { + Data []byte + Done bool +} + +// SSEReader 用于读取 SSE 流 +type SSEReader struct { + scanner *bufio.Scanner +} + +// NewSSEReader 创建 SSE 读取器 +func NewSSEReader(reader io.Reader) *SSEReader { + scanner := bufio.NewScanner(reader) + // 设置更大的缓冲区以处理长行(思维链内容可能很长) + buf := make([]byte, 1024*1024) + scanner.Buffer(buf, 1024*1024) + return &SSEReader{scanner: scanner} +} + +// ReadEvent 读取下一个 SSE 事件 +func (r *SSEReader) ReadEvent() (*SSEEvent, error) { + for r.scanner.Scan() { + line := r.scanner.Text() + + // 空行,跳过 + if line == "" { + continue + } + + // 检查是否为结束标记 + if line == "data: [DONE]" { + return &SSEEvent{Done: true}, nil + } + + // 解析 data 行 + if strings.HasPrefix(line, "data: ") { + jsonStr := line[6:] + return &SSEEvent{Data: []byte(jsonStr)}, nil + } + + // 其他行(如 event:, id: 等)跳过 + } + + if err := r.scanner.Err(); err != nil { + return nil, err + } + + return nil, errors.New("EOF") +} diff --git a/internal/models/provider/lkeap.go b/internal/models/provider/lkeap.go new file mode 100644 index 00000000..a3c6f963 --- /dev/null +++ b/internal/models/provider/lkeap.go @@ -0,0 +1,65 @@ +package provider + +import ( + "fmt" + "strings" + + "github.com/Tencent/WeKnora/internal/types" +) + +const ( + // LKEAPBaseURL 腾讯云知识引擎原子能力 (LKEAP) 兼容 OpenAI 协议的 BaseURL + LKEAPBaseURL = "https://api.lkeap.cloud.tencent.com/v1" +) + +// LKEAPProvider 实现腾讯云 LKEAP 的 Provider 接口 +// 支持 DeepSeek-R1, DeepSeek-V3 系列模型,具备思维链能力 +type LKEAPProvider struct{} + +func init() { + Register(&LKEAPProvider{}) +} + +// Info 返回 LKEAP provider 的元数据 +func (p *LKEAPProvider) Info() ProviderInfo { + return ProviderInfo{ + Name: ProviderLKEAP, + DisplayName: "腾讯云 LKEAP", + Description: "DeepSeek-R1, DeepSeek-V3 系列模型,支持思维链", + DefaultURLs: map[types.ModelType]string{ + types.ModelTypeKnowledgeQA: LKEAPBaseURL, + }, + ModelTypes: []types.ModelType{ + types.ModelTypeKnowledgeQA, + }, + RequiresAuth: true, + } +} + +// ValidateConfig 验证 LKEAP provider 配置 +func (p *LKEAPProvider) ValidateConfig(config *Config) error { + if config.APIKey == "" { + return fmt.Errorf("API key is required for LKEAP provider") + } + if config.ModelName == "" { + return fmt.Errorf("model name is required") + } + return nil +} + +// IsLKEAPDeepSeekV3Model 检查是否为 DeepSeek V3.x 系列模型 +// V3.x 系列支持通过 Thinking 参数控制思维链开关 +func IsLKEAPDeepSeekV3Model(modelName string) bool { + return strings.Contains(strings.ToLower(modelName), "deepseek-v3") +} + +// IsLKEAPDeepSeekR1Model 检查是否为 DeepSeek R1 系列模型 +// R1 系列默认开启思维链 +func IsLKEAPDeepSeekR1Model(modelName string) bool { + return strings.Contains(strings.ToLower(modelName), "deepseek-r1") +} + +// IsLKEAPThinkingModel 检查是否为支持思维链的 LKEAP 模型 +func IsLKEAPThinkingModel(modelName string) bool { + return IsLKEAPDeepSeekR1Model(modelName) || IsLKEAPDeepSeekV3Model(modelName) +} diff --git a/internal/models/provider/provider.go b/internal/models/provider/provider.go index 9742ac07..aabfb8f6 100644 --- a/internal/models/provider/provider.go +++ b/internal/models/provider/provider.go @@ -51,6 +51,8 @@ const ( ProviderQiniu ProviderName = "qiniu" // 美团 LongCat AI ProviderLongCat ProviderName = "longcat" + // 腾讯云 LKEAP (知识引擎原子能力) + ProviderLKEAP ProviderName = "lkeap" ) // AllProviders 返回所有注册的提供者名称 @@ -74,6 +76,7 @@ func AllProviders() []ProviderName { ProviderJina, ProviderMimo, ProviderLongCat, + ProviderLKEAP, ProviderGPUStack, } } @@ -238,6 +241,8 @@ func DetectProvider(baseURL string) ProviderName { return ProviderQianfan case containsAny(baseURL, "longcat.chat"): return ProviderLongCat + case containsAny(baseURL, "lkeap.cloud.tencent.com", "api.lkeap"): + return ProviderLKEAP default: return ProviderGeneric } diff --git a/internal/types/agent.go b/internal/types/agent.go index 203559f8..459cf75e 100644 --- a/internal/types/agent.go +++ b/internal/types/agent.go @@ -29,6 +29,8 @@ type AgentConfig struct { // MCP service selection MCPSelectionMode string `json:"mcp_selection_mode"` // MCP selection mode: "all", "selected", "none" MCPServices []string `json:"mcp_services"` // Selected MCP service IDs (when mode is "selected") + // Whether to enable thinking mode (for models that support extended thinking) + Thinking *bool `json:"thinking"` } // SessionAgentConfig represents session-level agent configuration diff --git a/internal/types/custom_agent.go b/internal/types/custom_agent.go index a74b739f..0d700a50 100644 --- a/internal/types/custom_agent.go +++ b/internal/types/custom_agent.go @@ -79,6 +79,8 @@ type CustomAgentConfig struct { Temperature float64 `yaml:"temperature" json:"temperature"` // Maximum completion tokens (only for normal mode) MaxCompletionTokens int `yaml:"max_completion_tokens" json:"max_completion_tokens"` + // Whether to enable thinking mode (for models that support extended thinking) + Thinking *bool `yaml:"thinking" json:"thinking"` // ===== Agent Mode Settings ===== // Maximum iterations for ReAct loop (only for agent type)