feat: 新增 AI 记账功能，包括流式聊天、洞察生成、语音转录和交易确认接口。
This commit is contained in:
@@ -1,6 +1,7 @@
|
||||
package service
|
||||
|
||||
import (
|
||||
"bufio"
|
||||
"bytes"
|
||||
"context"
|
||||
"crypto/md5"
|
||||
@@ -430,6 +431,17 @@ type ChatCompletionResponse struct {
|
||||
} `json:"choices"`
|
||||
}
|
||||
|
||||
// ChatCompletionStreamResponse represents OpenAI chat completion stream response
|
||||
type ChatCompletionStreamResponse struct {
|
||||
Choices []struct {
|
||||
Delta struct {
|
||||
Role string `json:"role,omitempty"`
|
||||
Content string `json:"content,omitempty"`
|
||||
} `json:"delta"`
|
||||
FinishReason string `json:"finish_reason,omitempty"`
|
||||
} `json:"choices"`
|
||||
}
|
||||
|
||||
// extractCustomPrompt 从用户消息中提取自定义 System/Persona prompt
|
||||
// 如果消息包含 "System:" 或 "Persona:"，则提取其后的内容作为自定义 prompt
|
||||
func extractCustomPrompt(text string) string {
|
||||
@@ -772,6 +784,99 @@ func (s *LLMService) GenerateReport(ctx context.Context, prompt string) (string,
|
||||
return chatResp.Choices[0].Message.Content, nil
|
||||
}
|
||||
|
||||
// StreamChat calls OpenAI Chat Completion API with streaming enabled
|
||||
// It returns a channel that yields text chunks, and an error if the request fails to start
|
||||
func (s *LLMService) StreamChat(ctx context.Context, messages []ChatMessage) (<-chan string, error) {
|
||||
if s.config.OpenAIAPIKey == "" || s.config.OpenAIBaseURL == "" {
|
||||
return nil, errors.New("OpenAI API not configured")
|
||||
}
|
||||
|
||||
reqBody := ChatCompletionRequest{
|
||||
Model: s.config.ChatModel,
|
||||
Messages: messages,
|
||||
Temperature: 0.1, // Lower temperature for consistent output in chat
|
||||
}
|
||||
|
||||
// Create a map to inject stream: true
|
||||
var bodyMap map[string]interface{}
|
||||
jsonBody, _ := json.Marshal(reqBody)
|
||||
json.Unmarshal(jsonBody, &bodyMap)
|
||||
bodyMap["stream"] = true
|
||||
|
||||
streamBody, err := json.Marshal(bodyMap)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("failed to marshal request: %w", err)
|
||||
}
|
||||
|
||||
req, err := http.NewRequestWithContext(ctx, "POST", s.config.OpenAIBaseURL+"/chat/completions", bytes.NewReader(streamBody))
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("failed to create request: %w", err)
|
||||
}
|
||||
|
||||
req.Header.Set("Authorization", "Bearer "+s.config.OpenAIAPIKey)
|
||||
req.Header.Set("Content-Type", "application/json")
|
||||
req.Header.Set("Accept", "text/event-stream")
|
||||
|
||||
// Use a separate client for streaming to avoid timeout issues?
|
||||
// Or just reuse. The standard timeout might be too short for long streams.
|
||||
// We'll trust the context cancellation for now.
|
||||
resp, err := s.httpClient.Do(req)
|
||||
if err != nil {
|
||||
return nil, fmt.Errorf("stream request failed: %w", err)
|
||||
}
|
||||
|
||||
if resp.StatusCode != http.StatusOK {
|
||||
resp.Body.Close()
|
||||
return nil, fmt.Errorf("stream failed with status %d", resp.StatusCode)
|
||||
}
|
||||
|
||||
outChan := make(chan string)
|
||||
|
||||
go func() {
|
||||
defer resp.Body.Close()
|
||||
defer close(outChan)
|
||||
|
||||
reader := bufio.NewReader(resp.Body)
|
||||
for {
|
||||
line, err := reader.ReadBytes('\n')
|
||||
if err != nil {
|
||||
if err != io.EOF {
|
||||
// Log error?
|
||||
}
|
||||
return
|
||||
}
|
||||
|
||||
line = bytes.TrimSpace(line)
|
||||
if !bytes.HasPrefix(line, []byte("data: ")) {
|
||||
continue
|
||||
}
|
||||
|
||||
data := bytes.TrimPrefix(line, []byte("data: "))
|
||||
if string(data) == "[DONE]" {
|
||||
return
|
||||
}
|
||||
|
||||
var streamResp ChatCompletionStreamResponse
|
||||
if err := json.Unmarshal(data, &streamResp); err != nil {
|
||||
continue
|
||||
}
|
||||
|
||||
if len(streamResp.Choices) > 0 {
|
||||
content := streamResp.Choices[0].Delta.Content
|
||||
if content != "" {
|
||||
select {
|
||||
case outChan <- content:
|
||||
case <-ctx.Done():
|
||||
return
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}()
|
||||
|
||||
return outChan, nil
|
||||
}
|
||||
|
||||
// MapAccountName maps natural language account name to account ID
|
||||
func (s *LLMService) MapAccountName(ctx context.Context, name string, userID uint) (*uint, string, error) {
|
||||
if name == "" {
|
||||
@@ -1033,6 +1138,194 @@ func (s *AIBookkeepingService) ProcessChat(ctx context.Context, userID uint, ses
|
||||
return response, nil
|
||||
}
|
||||
|
||||
// StreamProcessChat processes a chat message and streams the AI response
|
||||
// It calls onChunk callback for every text chunk received
|
||||
// Returns the final AIChatResponse after stream completion for metadata handling
|
||||
func (s *AIBookkeepingService) StreamProcessChat(ctx context.Context, userID uint, sessionID string, message string, onChunk func(string)) (*AIChatResponse, error) {
|
||||
session := s.getOrCreateSession(sessionID, userID)
|
||||
|
||||
// Add user message to history
|
||||
session.Messages = append(session.Messages, ChatMessage{
|
||||
Role: "user",
|
||||
Content: message,
|
||||
})
|
||||
|
||||
// 1. 获取财务上下文
|
||||
fc, _ := s.GetUserFinancialContext(ctx, userID)
|
||||
|
||||
// 2. 检测纯查询意图
|
||||
queryIntent := s.detectQueryIntent(message)
|
||||
if queryIntent != "" && fc != nil {
|
||||
// 查询意图通常较短，直接生成并模拟流式输出
|
||||
responseMsg := s.handleQueryIntent(ctx, queryIntent, message, fc)
|
||||
|
||||
// 模拟流式输出
|
||||
chunkSize := 4
|
||||
runes := []rune(responseMsg)
|
||||
for i := 0; i < len(runes); i += chunkSize {
|
||||
end := i + chunkSize
|
||||
if end > len(runes) {
|
||||
end = len(runes)
|
||||
}
|
||||
onChunk(string(runes[i:end]))
|
||||
time.Sleep(20 * time.Millisecond) // 稍微延迟模拟打字
|
||||
}
|
||||
|
||||
response := &AIChatResponse{
|
||||
SessionID: session.ID,
|
||||
Message: responseMsg,
|
||||
Intent: queryIntent,
|
||||
Params: session.Params,
|
||||
}
|
||||
|
||||
session.Messages = append(session.Messages, ChatMessage{
|
||||
Role: "assistant",
|
||||
Content: response.Message,
|
||||
})
|
||||
return response, nil
|
||||
}
|
||||
|
||||
// 3. 检测消费建议意图
|
||||
isSpendingAdvice := s.isSpendingAdviceIntent(message)
|
||||
|
||||
// Determine prompt
|
||||
var systemPrompt string
|
||||
// Fetch user categories for prompt context
|
||||
var categoryNamesStr string
|
||||
if categories, err := s.categoryRepo.GetAll(userID); err == nil && len(categories) > 0 {
|
||||
var names []string
|
||||
for _, c := range categories {
|
||||
names = append(names, c.Name)
|
||||
}
|
||||
categoryNamesStr = strings.Join(names, "/")
|
||||
}
|
||||
|
||||
todayDate := time.Now().Format("2006-01-02")
|
||||
catPrompt := "2. 分类：根据内容推断"
|
||||
if categoryNamesStr != "" {
|
||||
catPrompt = fmt.Sprintf("2. 分类：必须从以下已有分类中选择最匹配的一项：[%s]。", categoryNamesStr)
|
||||
}
|
||||
|
||||
systemPrompt = fmt.Sprintf(`你是一个智能记账助手。请以自然的对话方式回复用户，同时在回复中确认识别到的记账信息。
|
||||
今天的日期是%s
|
||||
|
||||
规则：
|
||||
1. 你的性格：贱萌、嘴硬心软、偶尔凡尔赛。
|
||||
2. 如果用户在记账：请提取金额、分类（%s）、描述。
|
||||
3. 如果用户在寻求建议：请给出毒舌但有用的建议。
|
||||
4. **不要返回 JSON**，直接以自然语言回复。`, todayDate, catPrompt)
|
||||
|
||||
// Construct minimal messages for stream
|
||||
streamMessages := []ChatMessage{
|
||||
{Role: "system", Content: systemPrompt},
|
||||
}
|
||||
// Add recent history
|
||||
historyLen := len(session.Messages)
|
||||
if historyLen > 4 {
|
||||
streamMessages = append(streamMessages, session.Messages[historyLen-4:]...)
|
||||
} else {
|
||||
streamMessages = append(streamMessages, session.Messages...)
|
||||
}
|
||||
|
||||
// Execute Stream
|
||||
streamChan, err := s.llmService.StreamChat(ctx, streamMessages)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
|
||||
var fullResponseBuilder strings.Builder
|
||||
for chunk := range streamChan {
|
||||
fullResponseBuilder.WriteString(chunk)
|
||||
onChunk(chunk)
|
||||
}
|
||||
fullResponse := fullResponseBuilder.String()
|
||||
|
||||
// 4. Update Session History with Assistant Message
|
||||
session.Messages = append(session.Messages, ChatMessage{
|
||||
Role: "assistant",
|
||||
Content: fullResponse,
|
||||
})
|
||||
|
||||
// 5. Post-Stream Logic: Parse Intent and Generate Card
|
||||
// We re-parse the *user message* (not the AI response) to get structured data
|
||||
// The original ParseIntent logic uses a specific JSON prompt which we skipped for the stream.
|
||||
// So we need to call ParseIntent again or have a way to extract it.
|
||||
|
||||
// OPTION A: Call ParseIntent strictly for data extraction (hidden from user)
|
||||
// This ensures we get the robust JSON parsing logic.
|
||||
params, _, err := s.llmService.ParseIntent(ctx, message, session.Messages[:len(session.Messages)-2], userID)
|
||||
if err != nil {
|
||||
// Log error but don't fail, user already saw the text
|
||||
fmt.Printf("Stream ParseIntent failed: %v", err)
|
||||
}
|
||||
|
||||
if params != nil {
|
||||
s.mergeParams(session.Params, params)
|
||||
}
|
||||
|
||||
// ... (Same mapping logic as ProcessChat)
|
||||
if session.Params.Account != "" && session.Params.AccountID == nil {
|
||||
accountID, accountName, _ := s.llmService.MapAccountName(ctx, session.Params.Account, userID)
|
||||
if accountID != nil {
|
||||
session.Params.AccountID = accountID
|
||||
session.Params.Account = accountName
|
||||
}
|
||||
}
|
||||
|
||||
if session.Params.Category != "" && session.Params.CategoryID == nil {
|
||||
categoryID, categoryName, _ := s.llmService.MapCategoryName(ctx, session.Params.Category, session.Params.Type, userID)
|
||||
if categoryID != nil {
|
||||
session.Params.CategoryID = categoryID
|
||||
session.Params.Category = categoryName
|
||||
}
|
||||
}
|
||||
|
||||
// Defaults...
|
||||
if session.Params.CategoryID == nil && session.Params.Category != "" {
|
||||
defaultCategoryID, defaultCategoryName := s.getDefaultCategory(userID, session.Params.Type)
|
||||
if defaultCategoryID != nil {
|
||||
session.Params.CategoryID = defaultCategoryID
|
||||
if session.Params.Category == "" {
|
||||
session.Params.Category = defaultCategoryName
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if session.Params.AccountID == nil {
|
||||
defaultAccountID, defaultAccountName := s.getDefaultAccount(userID, session.Params.Type)
|
||||
if defaultAccountID != nil {
|
||||
session.Params.AccountID = defaultAccountID
|
||||
session.Params.Account = defaultAccountName
|
||||
}
|
||||
}
|
||||
|
||||
// Construct final response object
|
||||
response := &AIChatResponse{
|
||||
SessionID: session.ID,
|
||||
Message: fullResponse, // The text already streamed
|
||||
Intent: "create_transaction",
|
||||
Params: session.Params,
|
||||
}
|
||||
|
||||
if isSpendingAdvice {
|
||||
response.Intent = "spending_advice"
|
||||
}
|
||||
|
||||
missingFields := s.getMissingFields(session.Params)
|
||||
if len(missingFields) > 0 {
|
||||
response.NeedsFollowUp = true
|
||||
// Note: We already streamed a response. We probably shouldn't override it with a follow-up question
|
||||
// unless the streamed response was generic.
|
||||
// For now, we trust the streamed response covers it, or the UI handles "NeedsFollowUp" silently if needed.
|
||||
// Actually, let's leave it to the UI. If NeedsFollowUp is true, maybe don't show confirmation card.
|
||||
} else {
|
||||
card := s.GenerateConfirmationCard(session)
|
||||
response.ConfirmationCard = card
|
||||
}
|
||||
|
||||
return response, nil
|
||||
}
|
||||
|
||||
// detectQueryIntent 检测用户查询意图
|
||||
func (s *AIBookkeepingService) detectQueryIntent(message string) string {
|
||||
budgetKeywords := []string{"预算", "剩多少", "还能花", "余额"} // 注意：余额可能指账户余额，这里简化处理
|
||||
|
||||
Reference in New Issue
Block a user