geekai/api/handler/chat_handler.go

package handler

// * +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
// * Copyright 2023 The Geek-AI Authors. All rights reserved.
// * Use of this source code is governed by a Apache-2.0 license
// * that can be found in the LICENSE file.
// * @Author yangjian102621@163.com
// * +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

import (
	"bufio"
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"geekai/core"
	"geekai/core/types"
	"geekai/service"
	"geekai/service/oss"
	"geekai/store/model"
	"geekai/store/vo"
	"geekai/utils"
	"geekai/utils/resp"
	"html/template"
	"io"
	"net/http"
	"net/url"
	"os"
	"strings"
	"time"
	"unicode/utf8"

	"github.com/gin-gonic/gin"
	"github.com/go-redis/redis/v8"
	req2 "github.com/imroc/req/v3"
	"github.com/sashabaranov/go-openai"
	"gorm.io/gorm"
)

const (
	ChatEventStart        = "start"
	ChatEventEnd          = "end"
	ChatEventError        = "error"
	ChatEventMessageDelta = "message_delta"
	ChatEventTitle        = "title"
)

type ChatInput struct {
	UserId  uint      `json:"user_id"`
	RoleId  int       `json:"role_id"`
	ModelId int       `json:"model_id"`
	ChatId  string    `json:"chat_id"`
	Content string    `json:"content"`
	Tools   []int     `json:"tools"`
	Stream  bool      `json:"stream"`
	Files   []vo.File `json:"files"`
}

type ChatHandler struct {
	BaseHandler
	redis          *redis.Client
	uploadManager  *oss.UploaderManager
	licenseService *service.LicenseService
	ReqCancelFunc  *types.LMap[string, context.CancelFunc] // HttpClient 请求取消 handle function
	ChatContexts   *types.LMap[string, []any]              // 聊天上下文 Map [chatId] => []Message
	userService    *service.UserService
}

func NewChatHandler(app *core.AppServer, db *gorm.DB, redis *redis.Client, manager *oss.UploaderManager, licenseService *service.LicenseService, userService *service.UserService) *ChatHandler {
	return &ChatHandler{
		BaseHandler:    BaseHandler{App: app, DB: db},
		redis:          redis,
		uploadManager:  manager,
		licenseService: licenseService,
		ReqCancelFunc:  types.NewLMap[string, context.CancelFunc](),
		ChatContexts:   types.NewLMap[string, []any](),
		userService:    userService,
	}
}

// Chat 处理聊天请求
func (h *ChatHandler) Chat(c *gin.Context) {
	var data ChatInput
	if err := c.ShouldBindJSON(&data); err != nil {
		resp.ERROR(c, types.InvalidArgs)
		return
	}

	// 设置SSE响应头
	c.Header("Content-Type", "text/event-stream")
	c.Header("Cache-Control", "no-cache")
	c.Header("Connection", "keep-alive")
	c.Header("X-Accel-Buffering", "no")

	ctx, cancel := context.WithCancel(c.Request.Context())
	defer cancel()

	// 验证聊天角色
	var chatRole model.ChatRole
	err := h.DB.First(&chatRole, data.RoleId).Error
	if err != nil || !chatRole.Enable {
		pushMessage(c, ChatEventError, "当前聊天角色不存在或者未启用，请更换角色之后再发起对话！")
		return
	}

	// 如果角色绑定了模型ID，使用角色的模型ID
	if chatRole.ModelId > 0 {
		data.ModelId = int(chatRole.ModelId)
	}

	// 获取模型信息
	var chatModel model.ChatModel
	err = h.DB.Where("id", data.ModelId).First(&chatModel).Error
	if err != nil || !chatModel.Enabled {
		pushMessage(c, ChatEventError, "当前AI模型暂未启用，请更换模型后再发起对话！")
		return
	}

	session := &types.ChatSession{
		ClientIP: c.ClientIP(),
		UserId:   data.UserId,
		ChatId:   data.ChatId,
		Tools:    data.Tools,
		Stream:   data.Stream,
		Model: types.ChatModel{
			KeyId: data.ModelId,
		},
	}

	// 使用旧的聊天数据覆盖模型和角色ID
	var chat model.ChatItem
	h.DB.Where("chat_id", data.ChatId).First(&chat)
	if chat.Id > 0 {
		chatModel.Id = chat.ModelId
		data.RoleId = int(chat.RoleId)
	}

	// 复制模型数据
	err = utils.CopyObject(chatModel, &session.Model)
	if err != nil {
		logger.Error(err, chatModel)
	}
	session.Model.Id = chatModel.Id

	// 发送消息
	err = h.sendMessage(ctx, session, chatRole, data.Content, c)
	if err != nil {
		pushMessage(c, ChatEventError, err.Error())
		return
	}

	pushMessage(c, ChatEventEnd, "对话完成")
}

func pushMessage(c *gin.Context, msgType string, content interface{}) {
	c.SSEvent("message", map[string]interface{}{
		"type": msgType,
		"body": content,
	})
	c.Writer.Flush()
}

func (h *ChatHandler) sendMessage(ctx context.Context, session *types.ChatSession, role model.ChatRole, prompt string, c *gin.Context) error {
	var user model.User
	res := h.DB.Model(&model.User{}).First(&user, session.UserId)
	if res.Error != nil {
		return errors.New("未授权用户，您正在进行非法操作！")
	}
	var userVo vo.User
	err := utils.CopyObject(user, &userVo)
	userVo.Id = user.Id
	if err != nil {
		return errors.New("User 对象转换失败，" + err.Error())
	}

	if !userVo.Status {
		return errors.New("您的账号已经被禁用，如果疑问，请联系管理员！")
	}

	if userVo.Power < session.Model.Power {
		return fmt.Errorf("您当前剩余算力 %d 已不足以支付当前模型的单次对话需要消耗的算力 %d，[立即购买](/member)。", userVo.Power, session.Model.Power)
	}

	if userVo.ExpiredTime > 0 && userVo.ExpiredTime <= time.Now().Unix() {
		return errors.New("您的账号已经过期，请联系管理员！")
	}

	// 检查 prompt 长度是否超过了当前模型允许的最大上下文长度
	promptTokens, _ := utils.CalcTokens(prompt, session.Model.Value)
	if promptTokens > session.Model.MaxContext {

		return errors.New("对话内容超出了当前模型允许的最大上下文长度！")
	}

	var req = types.ApiRequest{
		Model:       session.Model.Value,
		Stream:      session.Stream,
		Temperature: session.Model.Temperature,
	}
	// 兼容 OpenAI 模型
	if strings.HasPrefix(session.Model.Value, "o1-") ||
		strings.HasPrefix(session.Model.Value, "o3-") ||
		strings.HasPrefix(session.Model.Value, "gpt") {
		req.MaxCompletionTokens = session.Model.MaxTokens
		session.Start = time.Now().Unix()
	} else {
		req.MaxTokens = session.Model.MaxTokens
	}

	if len(session.Tools) > 0 && !strings.HasPrefix(session.Model.Value, "o1-") {
		var items []model.Function
		res = h.DB.Where("enabled", true).Where("id IN ?", session.Tools).Find(&items)
		if res.Error == nil {
			var tools = make([]types.Tool, 0)
			for _, v := range items {
				var parameters map[string]interface{}
				err = utils.JsonDecode(v.Parameters, &parameters)
				if err != nil {
					continue
				}
				tool := types.Tool{
					Type: "function",
					Function: types.Function{
						Name:        v.Name,
						Description: v.Description,
						Parameters:  parameters,
					},
				}
				if v, ok := parameters["required"]; v == nil || !ok {
					tool.Function.Parameters["required"] = []string{}
				}
				tools = append(tools, tool)
			}

			if len(tools) > 0 {
				req.Tools = tools
				req.ToolChoice = "auto"
			}
		}
	}

	// 加载聊天上下文
	chatCtx := make([]interface{}, 0)
	messages := make([]interface{}, 0)
	if h.App.SysConfig.EnableContext {
		if h.ChatContexts.Has(session.ChatId) {
			messages = h.ChatContexts.Get(session.ChatId)
		} else {
			_ = utils.JsonDecode(role.Context, &messages)
			if h.App.SysConfig.ContextDeep > 0 {
				var historyMessages []model.ChatMessage
				res := h.DB.Where("chat_id = ? and use_context = 1", session.ChatId).Limit(h.App.SysConfig.ContextDeep).Order("id DESC").Find(&historyMessages)
				if res.Error == nil {
					for i := len(historyMessages) - 1; i >= 0; i-- {
						msg := historyMessages[i]
						ms := types.Message{Role: "user", Content: msg.Content}
						if msg.Type == types.ReplyMsg {
							ms.Role = "assistant"
						}
						chatCtx = append(chatCtx, ms)
					}
				}
			}
		}

		// 计算当前请求的 token 总长度，确保不会超出最大上下文长度
		// MaxContextLength = Response + Tool + Prompt + Context
		tokens := req.MaxTokens // 最大响应长度
		tks, _ := utils.CalcTokens(utils.JsonEncode(req.Tools), req.Model)
		tokens += tks + promptTokens

		for i := len(messages) - 1; i >= 0; i-- {
			v := messages[i]
			tks, _ = utils.CalcTokens(utils.JsonEncode(v), req.Model)
			// 上下文 token 超出了模型的最大上下文长度
			if tokens+tks >= session.Model.MaxContext {
				break
			}

			// 上下文的深度超出了模型的最大上下文深度
			if len(chatCtx) >= h.App.SysConfig.ContextDeep {
				break
			}

			tokens += tks
			chatCtx = append(chatCtx, v)
		}

		logger.Debugf("聊天上下文：%+v", chatCtx)
	}
	reqMgs := make([]interface{}, 0)

	for i := len(chatCtx) - 1; i >= 0; i-- {
		reqMgs = append(reqMgs, chatCtx[i])
	}

	fullPrompt := prompt
	text := prompt
	// extract files in prompt
	files := utils.ExtractFileURLs(prompt)
	logger.Debugf("detected FILES: %+v", files)
	// 如果不是逆向模型，则提取文件内容
	if len(files) > 0 && !(session.Model.Value == "gpt-4-all" ||
		strings.HasPrefix(session.Model.Value, "gpt-4-gizmo") ||
		strings.HasPrefix(session.Model.Value, "claude-3")) {
		contents := make([]string, 0)
		var file model.File
		for _, v := range files {
			h.DB.Where("url = ?", v).First(&file)
			content, err := utils.ReadFileContent(v, h.App.Config.TikaHost)
			if err != nil {
				logger.Error("error with read file: ", err)
			} else {
				contents = append(contents, fmt.Sprintf("%s 文件内容：%s", file.Name, content))
			}
			text = strings.Replace(text, v, "", 1)
		}
		if len(contents) > 0 {
			fullPrompt = fmt.Sprintf("请根据提供的文件内容信息回答问题(其中Excel 已转成 HTML)：\n\n %s\n\n 问题：%s", strings.Join(contents, "\n"), text)
		}

		tokens, _ := utils.CalcTokens(fullPrompt, req.Model)
		if tokens > session.Model.MaxContext {
			return fmt.Errorf("文件的长度超出模型允许的最大上下文长度，请减少文件内容数量或文件大小。")
		}
	}
	logger.Debug("最终Prompt：", fullPrompt)

	// extract images from prompt
	imgURLs := utils.ExtractImgURLs(prompt)
	logger.Debugf("detected IMG: %+v", imgURLs)
	var content interface{}
	if len(imgURLs) > 0 {
		data := make([]interface{}, 0)
		for _, v := range imgURLs {
			text = strings.Replace(text, v, "", 1)
			data = append(data, gin.H{
				"type": "image_url",
				"image_url": gin.H{
					"url": v,
				},
			})
		}
		data = append(data, gin.H{
			"type": "text",
			"text": strings.TrimSpace(text),
		})
		content = data
	} else {
		content = fullPrompt
	}
	req.Messages = append(reqMgs, map[string]interface{}{
		"role":    "user",
		"content": content,
	})

	logger.Debugf("%+v", req.Messages)

	return h.sendOpenAiMessage(req, userVo, ctx, session, role, prompt, c)
}

// Tokens 统计 token 数量
func (h *ChatHandler) Tokens(c *gin.Context) {
	var data struct {
		Text   string `json:"text"`
		Model  string `json:"model"`
		ChatId string `json:"chat_id"`
	}
	if err := c.ShouldBindJSON(&data); err != nil {
		resp.ERROR(c, types.InvalidArgs)
		return
	}

	// 如果没有传入 text 字段，则说明是获取当前 reply 总的 token 消耗（带上下文）
	if data.Text == "" && data.ChatId != "" {
		var item model.ChatMessage
		userId, _ := c.Get(types.LoginUserID)
		res := h.DB.Where("user_id = ?", userId).Where("chat_id = ?", data.ChatId).Last(&item)
		if res.Error != nil {
			resp.ERROR(c, res.Error.Error())
			return
		}
		resp.SUCCESS(c, item.Tokens)
		return
	}

	tokens, err := utils.CalcTokens(data.Text, data.Model)
	if err != nil {
		resp.ERROR(c, err.Error())
		return
	}

	resp.SUCCESS(c, tokens)
}

func getTotalTokens(req types.ApiRequest) int {
	encode := utils.JsonEncode(req.Messages)
	var items []map[string]interface{}
	err := utils.JsonDecode(encode, &items)
	if err != nil {
		return 0
	}
	tokens := 0
	for _, item := range items {
		content, ok := item["content"]
		if ok && !utils.IsEmptyValue(content) {
			t, err := utils.CalcTokens(utils.InterfaceToString(content), req.Model)
			if err == nil {
				tokens += t
			}
		}
	}
	return tokens
}

// StopGenerate 停止生成
func (h *ChatHandler) StopGenerate(c *gin.Context) {
	sessionId := c.Query("session_id")
	if h.ReqCancelFunc.Has(sessionId) {
		h.ReqCancelFunc.Get(sessionId)()
		h.ReqCancelFunc.Delete(sessionId)
	}
	resp.SUCCESS(c, types.OkMsg)
}

// 发送请求到 OpenAI 服务器
// useOwnApiKey: 是否使用了用户自己的 API KEY
func (h *ChatHandler) doRequest(ctx context.Context, req types.ApiRequest, session *types.ChatSession, apiKey *model.ApiKey) (*http.Response, error) {
	// if the chat model bind a KEY, use it directly
	if session.Model.KeyId > 0 {
		h.DB.Where("id", session.Model.KeyId).Find(apiKey)
	} else { // use the last unused key
		h.DB.Where("type", "chat").Where("enabled", true).Order("last_used_at ASC").First(apiKey)
	}

	if apiKey.Id == 0 {
		return nil, errors.New("no available key, please import key")
	}

	// ONLY allow apiURL in blank list
	err := h.licenseService.IsValidApiURL(apiKey.ApiURL)
	if err != nil {
		return nil, err
	}
	logger.Debugf("对话请求消息体：%+v", req)
	var apiURL string
	p, _ := url.Parse(apiKey.ApiURL)
	// 如果设置的是 BASE_URL 没有路径，则添加 /v1/chat/completions
	if p.Path == "" {
		apiURL = fmt.Sprintf("%s/v1/chat/completions", apiKey.ApiURL)
	} else {
		apiURL = apiKey.ApiURL
	}
	// 创建 HttpClient 请求对象
	var client *http.Client
	requestBody, err := json.Marshal(req)
	if err != nil {
		return nil, err
	}
	request, err := http.NewRequest(http.MethodPost, apiURL, bytes.NewBuffer(requestBody))
	if err != nil {
		return nil, err
	}

	request = request.WithContext(ctx)
	request.Header.Set("Content-Type", "application/json")
	if len(apiKey.ProxyURL) > 5 { // 使用代理
		proxy, _ := url.Parse(apiKey.ProxyURL)
		client = &http.Client{
			Transport: &http.Transport{
				Proxy: http.ProxyURL(proxy),
			},
		}
	} else {
		client = http.DefaultClient
	}
	logger.Infof("Sending %s request, API KEY:%s, PROXY: %s, Model: %s", apiKey.ApiURL, apiURL, apiKey.ProxyURL, req.Model)
	request.Header.Set("Authorization", fmt.Sprintf("Bearer %s", apiKey.Value))
	// 更新API KEY 最后使用时间
	h.DB.Model(&model.ApiKey{}).Where("id", apiKey.Id).UpdateColumn("last_used_at", time.Now().Unix())
	return client.Do(request)
}

// 扣减用户算力
func (h *ChatHandler) subUserPower(userVo vo.User, session *types.ChatSession, promptTokens int, replyTokens int) {
	power := 1
	if session.Model.Power > 0 {
		power = session.Model.Power
	}

	err := h.userService.DecreasePower(userVo.Id, power, model.PowerLog{
		Type:   types.PowerConsume,
		Model:  session.Model.Value,
		Remark: fmt.Sprintf("模型名称：%s, 提问长度：%d，回复长度：%d", session.Model.Name, promptTokens, replyTokens),
	})
	if err != nil {
		logger.Error(err)
	}
}

func (h *ChatHandler) saveChatHistory(
	req types.ApiRequest,
	usage Usage,
	message types.Message,
	session *types.ChatSession,
	role model.ChatRole,
	userVo vo.User,
	promptCreatedAt time.Time,
	replyCreatedAt time.Time) {

	// 更新上下文消息
	if h.App.SysConfig.EnableContext {
		chatCtx := req.Messages            // 提问消息
		chatCtx = append(chatCtx, message) // 回复消息
		h.ChatContexts.Put(session.ChatId, chatCtx)
	}

	// 追加聊天记录
	// for prompt
	var promptTokens, replyTokens, totalTokens int
	if usage.PromptTokens > 0 {
		promptTokens = usage.PromptTokens
	} else {
		promptTokens, _ = utils.CalcTokens(usage.Content, req.Model)
	}

	historyUserMsg := model.ChatMessage{
		UserId:      userVo.Id,
		ChatId:      session.ChatId,
		RoleId:      role.Id,
		Type:        types.PromptMsg,
		Icon:        userVo.Avatar,
		Content:     template.HTMLEscapeString(usage.Prompt),
		Tokens:      promptTokens,
		TotalTokens: promptTokens,
		UseContext:  true,
		Model:       req.Model,
	}
	historyUserMsg.CreatedAt = promptCreatedAt
	historyUserMsg.UpdatedAt = promptCreatedAt
	err := h.DB.Save(&historyUserMsg).Error
	if err != nil {
		logger.Error("failed to save prompt history message: ", err)
	}

	// for reply
	// 计算本次对话消耗的总 token 数量
	if usage.CompletionTokens > 0 {
		replyTokens = usage.CompletionTokens
		totalTokens = usage.TotalTokens
	} else {
		replyTokens, _ = utils.CalcTokens(message.Content, req.Model)
		totalTokens = replyTokens + getTotalTokens(req)
	}
	historyReplyMsg := model.ChatMessage{
		UserId:      userVo.Id,
		ChatId:      session.ChatId,
		RoleId:      role.Id,
		Type:        types.ReplyMsg,
		Icon:        role.Icon,
		Content:     usage.Content,
		Tokens:      replyTokens,
		TotalTokens: totalTokens,
		UseContext:  true,
		Model:       req.Model,
	}
	historyReplyMsg.CreatedAt = replyCreatedAt
	historyReplyMsg.UpdatedAt = replyCreatedAt
	err = h.DB.Create(&historyReplyMsg).Error
	if err != nil {
		logger.Error("failed to save reply history message: ", err)
	}

	// 更新用户算力
	if session.Model.Power > 0 {
		h.subUserPower(userVo, session, promptTokens, replyTokens)
	}
	// 保存当前会话
	var chatItem model.ChatItem
	err = h.DB.Where("chat_id = ?", session.ChatId).First(&chatItem).Error
	if err != nil {
		chatItem.ChatId = session.ChatId
		chatItem.UserId = userVo.Id
		chatItem.RoleId = role.Id
		chatItem.ModelId = session.Model.Id
		if utf8.RuneCountInString(usage.Prompt) > 30 {
			chatItem.Title = string([]rune(usage.Prompt)[:30]) + "..."
		} else {
			chatItem.Title = usage.Prompt
		}
		chatItem.Model = req.Model
		err = h.DB.Create(&chatItem).Error
		if err != nil {
			logger.Error("failed to save chat item: ", err)
		}
	}
}

// 文本生成语音
func (h *ChatHandler) TextToSpeech(c *gin.Context) {
	var data struct {
		ModelId int    `json:"model_id"`
		Text    string `json:"text"`
	}
	if err := c.ShouldBindJSON(&data); err != nil {
		resp.ERROR(c, types.InvalidArgs)
		return
	}

	textHash := utils.Sha256(fmt.Sprintf("%d/%s", data.ModelId, data.Text))
	audioFile := fmt.Sprintf("%s/audio", h.App.Config.StaticDir)
	if _, err := os.Stat(audioFile); err != nil {
		os.MkdirAll(audioFile, 0755)
	}
	audioFile = fmt.Sprintf("%s/%s.mp3", audioFile, textHash)
	if _, err := os.Stat(audioFile); err == nil {
		// 设置响应头
		c.Header("Content-Type", "audio/mpeg")
		c.Header("Content-Disposition", "attachment; filename=speech.mp3")
		c.File(audioFile)
		return
	}

	// 查询模型
	var chatModel model.ChatModel
	err := h.DB.Where("id", data.ModelId).First(&chatModel).Error
	if err != nil {
		resp.ERROR(c, "找不到语音模型")
		return
	}

	// 调用 DeepSeek 的 API 接口
	var apiKey model.ApiKey
	if chatModel.KeyId > 0 {
		h.DB.Where("id", chatModel.KeyId).First(&apiKey)
	}
	if apiKey.Id == 0 {
		h.DB.Where("type", "tts").Where("enabled", true).First(&apiKey)
	}
	if apiKey.Id == 0 {
		resp.ERROR(c, "no TTS API key, please import key")
		return
	}

	logger.Debugf("chatModel: %+v, apiKey: %+v", chatModel, apiKey)

	// 调用 openai tts api
	config := openai.DefaultConfig(apiKey.Value)
	config.BaseURL = apiKey.ApiURL + "/v1"
	client := openai.NewClientWithConfig(config)
	voice := openai.VoiceAlloy
	var options map[string]string
	err = utils.JsonDecode(chatModel.Options, &options)
	if err == nil {
		voice = openai.SpeechVoice(options["voice"])
	}
	req := openai.CreateSpeechRequest{
		Model: openai.SpeechModel(chatModel.Value),
		Input: data.Text,
		Voice: voice,
	}

	audioData, err := client.CreateSpeech(context.Background(), req)
	if err != nil {
		resp.ERROR(c, err.Error())
		return
	}

	// 先将音频数据读取到内存
	audioBytes, err := io.ReadAll(audioData)
	if err != nil {
		resp.ERROR(c, err.Error())
		return
	}

	// 保存到音频文件
	err = os.WriteFile(audioFile, audioBytes, 0644)
	if err != nil {
		logger.Error("failed to save audio file: ", err)
	}

	// 设置响应头
	c.Header("Content-Type", "audio/mpeg")
	c.Header("Content-Disposition", "attachment; filename=speech.mp3")

	// 直接写入完整的音频数据到响应
	c.Writer.Write(audioBytes)
}

// OPenAI 消息发送实现
func (h *ChatHandler) sendOpenAiMessage(
	req types.ApiRequest,
	userVo vo.User,
	ctx context.Context,
	session *types.ChatSession,
	role model.ChatRole,
	prompt string,
	c *gin.Context) error {
	promptCreatedAt := time.Now() // 记录提问时间
	start := time.Now()
	var apiKey = model.ApiKey{}
	response, err := h.doRequest(ctx, req, session, &apiKey)
	logger.Info("HTTP请求完成，耗时：", time.Since(start))
	if err != nil {
		if strings.Contains(err.Error(), "context canceled") {
			return fmt.Errorf("用户取消了请求：%s", prompt)
		} else if strings.Contains(err.Error(), "no available key") {
			return errors.New("抱歉😔😔😔，系统已经没有可用的 API KEY，请联系管理员！")
		}
		return err
	} else {
		defer response.Body.Close()
	}

	if response.StatusCode != 200 {
		body, _ := io.ReadAll(response.Body)
		return fmt.Errorf("请求 OpenAI API 失败：%d, %v", response.StatusCode, string(body))
	}

	contentType := response.Header.Get("Content-Type")
	if strings.Contains(contentType, "text/event-stream") {
		replyCreatedAt := time.Now() // 记录回复时间
		// 循环读取 Chunk 消息
		var message = types.Message{Role: "assistant"}
		var contents = make([]string, 0)
		var function model.Function
		var toolCall = false
		var arguments = make([]string, 0)
		var reasoning = false

		pushMessage(c, ChatEventStart, "开始响应")
		scanner := bufio.NewScanner(response.Body)
		for scanner.Scan() {
			line := scanner.Text()
			if !strings.Contains(line, "data:") || len(line) < 30 {
				continue
			}
			var responseBody = types.ApiResponse{}
			err = json.Unmarshal([]byte(line[6:]), &responseBody)
			if err != nil { // 数据解析出错
				return errors.New(line)
			}
			if len(responseBody.Choices) == 0 { // Fixed: 兼容 Azure API 第一个输出空行
				continue
			}
			if responseBody.Choices[0].Delta.Content == nil &&
				responseBody.Choices[0].Delta.ToolCalls == nil &&
				responseBody.Choices[0].Delta.ReasoningContent == "" {
				continue
			}

			if responseBody.Choices[0].FinishReason == "stop" && len(contents) == 0 {
				pushMessage(c, ChatEventError, "抱歉😔😔😔，AI助手由于未知原因已经停止输出内容。")
				break
			}

			var tool types.ToolCall
			if len(responseBody.Choices[0].Delta.ToolCalls) > 0 {
				tool = responseBody.Choices[0].Delta.ToolCalls[0]
				if toolCall && tool.Function.Name == "" {
					arguments = append(arguments, tool.Function.Arguments)
					continue
				}
			}

			// 兼容 Function Call
			fun := responseBody.Choices[0].Delta.FunctionCall
			if fun.Name != "" {
				tool = *new(types.ToolCall)
				tool.Function.Name = fun.Name
			} else if toolCall {
				arguments = append(arguments, fun.Arguments)
				continue
			}

			if !utils.IsEmptyValue(tool) {
				res := h.DB.Where("name = ?", tool.Function.Name).First(&function)
				if res.Error == nil {
					toolCall = true
					callMsg := fmt.Sprintf("正在调用工具 `%s` 作答 ...\n\n", function.Label)
					pushMessage(c, ChatEventMessageDelta, map[string]interface{}{
						"type":    "text",
						"content": callMsg,
					})
					contents = append(contents, callMsg)
				}
				continue
			}

			if responseBody.Choices[0].FinishReason == "tool_calls" ||
				responseBody.Choices[0].FinishReason == "function_call" { // 函数调用完毕
				break
			}

			// output stopped
			if responseBody.Choices[0].FinishReason != "" {
				break // 输出完成或者输出中断了
			} else { // 正常输出结果
				// 兼容思考过程
				if responseBody.Choices[0].Delta.ReasoningContent != "" {
					reasoningContent := responseBody.Choices[0].Delta.ReasoningContent
					if !reasoning {
						reasoningContent = fmt.Sprintf("<think>%s", reasoningContent)
						reasoning = true
					}

					pushMessage(c, ChatEventMessageDelta, map[string]interface{}{
						"type":    "text",
						"content": reasoningContent,
					})
					contents = append(contents, reasoningContent)
				} else if responseBody.Choices[0].Delta.Content != "" {
					finalContent := responseBody.Choices[0].Delta.Content
					if reasoning {
						finalContent = fmt.Sprintf("</think>%s", responseBody.Choices[0].Delta.Content)
						reasoning = false
					}
					contents = append(contents, utils.InterfaceToString(finalContent))
					pushMessage(c, ChatEventMessageDelta, map[string]interface{}{
						"type":    "text",
						"content": finalContent,
					})
				}
			}
		} // end for

		if err := scanner.Err(); err != nil {
			if strings.Contains(err.Error(), "context canceled") {
				logger.Info("用户取消了请求：", prompt)
			} else {
				logger.Error("信息读取出错：", err)
			}
		}

		if toolCall { // 调用函数完成任务
			params := make(map[string]any)
			_ = utils.JsonDecode(strings.Join(arguments, ""), &params)
			logger.Debugf("函数名称: %s, 函数参数：%s", function.Name, params)
			params["user_id"] = userVo.Id
			var apiRes types.BizVo
			r, err := req2.C().R().SetHeader("Body-Type", "application/json").
				SetHeader("Authorization", function.Token).
				SetBody(params).Post(function.Action)
			errMsg := ""
			if err != nil {
				errMsg = err.Error()
			} else {
				all, _ := io.ReadAll(r.Body)
				err = json.Unmarshal(all, &apiRes)
				if err != nil {
					errMsg = err.Error()
				} else if apiRes.Code != types.Success {
					errMsg = apiRes.Message
				}
			}

			if errMsg != "" {
				errMsg = "调用函数工具出错：" + errMsg
				contents = append(contents, errMsg)
			} else {
				errMsg = utils.InterfaceToString(apiRes.Data)
				contents = append(contents, errMsg)
			}
			pushMessage(c, ChatEventMessageDelta, map[string]interface{}{
				"type":    "text",
				"content": errMsg,
			})
		}

		// 消息发送成功
		if len(contents) > 0 {
			usage := Usage{
				Prompt:           prompt,
				Content:          strings.Join(contents, ""),
				PromptTokens:     0,
				CompletionTokens: 0,
				TotalTokens:      0,
			}
			message.Content = usage.Content
			h.saveChatHistory(req, usage, message, session, role, userVo, promptCreatedAt, replyCreatedAt)
		}
	} else {
		var respVo OpenAIResVo
		body, err := io.ReadAll(response.Body)
		if err != nil {
			return fmt.Errorf("读取响应失败：%v", body)
		}
		err = json.Unmarshal(body, &respVo)
		if err != nil {
			return fmt.Errorf("解析响应失败：%v", body)
		}
		content := respVo.Choices[0].Message.Content
		if strings.HasPrefix(req.Model, "o1-") {
			content = fmt.Sprintf("AI思考结束，耗时：%d 秒。\n%s", time.Now().Unix()-session.Start, respVo.Choices[0].Message.Content)
		}
		pushMessage(c, ChatEventMessageDelta, map[string]interface{}{
			"type":    "text",
			"content": content,
		})
		respVo.Usage.Prompt = prompt
		respVo.Usage.Content = content
		h.saveChatHistory(req, respVo.Usage, respVo.Choices[0].Message, session, role, userVo, promptCreatedAt, time.Now())
	}

	return nil
}