feat: 实现 OpenAI LLM 服务（SSE 流式解析 + 多模态消息构建）及 System Prompt 定义

backend/internal/ai/llm/openai.go

@@ -0,0 +1,237 @@
+package llm
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"go.uber.org/zap"
+)
+
+// OpenAIService 基于 OpenAI Chat Completions API 的 LLM 实现。
+type OpenAIService struct {
+	apiKey   string
+	model    string
+	endpoint string
+	timeout  time.Duration
+	logger   *zap.SugaredLogger
+	client   *http.Client
+}
+
+// NewOpenAIService 创建 OpenAI LLM 服务。
+func NewOpenAIService(apiKey, model, endpoint string, timeoutSec int, logger *zap.SugaredLogger) *OpenAIService {
+	if model == "" {
+		model = "gpt-4o"
+	}
+	if endpoint == "" {
+		endpoint = "https://api.openai.com/v1"
+	}
+	timeout := time.Duration(timeoutSec) * time.Second
+	if timeout <= 0 {
+		timeout = 10 * time.Second
+	}
+	return &OpenAIService{
+		apiKey:   apiKey,
+		model:    model,
+		endpoint: endpoint,
+		timeout:  timeout,
+		logger:   logger,
+		client:   &http.Client{Timeout: 60 * time.Second}, // HTTP client timeout > LLM timeout
+	}
+}
+
+// --- OpenAI API 请求/响应结构 ---
+
+type chatRequest struct {
+	Model    string        `json:"model"`
+	Messages []chatMessage `json:"messages"`
+	Stream   bool          `json:"stream"`
+}
+
+type chatMessage struct {
+	Role    string        `json:"role"`
+	Content []contentPart `json:"content"`
+}
+
+type contentPart struct {
+	Type     string    `json:"type"`
+	Text     string    `json:"text,omitempty"`
+	ImageURL *imageURL `json:"image_url,omitempty"`
+}
+
+type imageURL struct {
+	URL string `json:"url"`
+}
+
+// streamDelta SSE 流式响应的单个 delta。
+type streamDelta struct {
+	Choices []struct {
+		Delta struct {
+			Content string `json:"content"`
+		} `json:"delta"`
+		FinishReason *string `json:"finish_reason"`
+	} `json:"choices"`
+	Usage *struct {
+		PromptTokens     int `json:"prompt_tokens"`
+		CompletionTokens int `json:"completion_tokens"`
+		TotalTokens      int `json:"total_tokens"`
+	} `json:"usage"`
+	Model string `json:"model"`
+}
+
+// ChatStream 实现 llm.Service。调用 OpenAI Chat Completions API 流式推理。
+func (o *OpenAIService) ChatStream(ctx context.Context, req Request) (<-chan Chunk, error) {
+	// 构建请求
+	messages := o.buildMessages(req)
+
+	body := chatRequest{
+		Model:    o.model,
+		Messages: messages,
+		Stream:   true,
+	}
+
+	payload, err := json.Marshal(body)
+	if err != nil {
+		return nil, fmt.Errorf("llm: marshal request: %w", err)
+	}
+
+	// 创建带超时的 context
+	ctx, cancel := context.WithTimeout(ctx, o.timeout)
+
+	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, o.endpoint+"/chat/completions", bytes.NewReader(payload))
+	if err != nil {
+		cancel()
+		return nil, fmt.Errorf("llm: create request: %w", err)
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	httpReq.Header.Set("Authorization", "Bearer "+o.apiKey)
+
+	resp, err := o.client.Do(httpReq)
+	if err != nil {
+		cancel()
+		return nil, fmt.Errorf("llm: send request: %w", err)
+	}
+
+	if resp.StatusCode != http.StatusOK {
+		cancel()
+		bodyBytes, _ := io.ReadAll(resp.Body)
+		resp.Body.Close()
+		return nil, fmt.Errorf("llm: api error (status %d): %s", resp.StatusCode, string(bodyBytes))
+	}
+
+	// 启动 goroutine 解析 SSE 流
+	ch := make(chan Chunk, 64)
+	go func() {
+		defer close(ch)
+		defer cancel()
+		defer resp.Body.Close()
+
+		o.parseSSEStream(resp.Body, ch)
+	}()
+
+	return ch, nil
+}
+
+// parseSSEStream 解析 SSE 流，将 delta 发送到 channel。
+func (o *OpenAIService) parseSSEStream(body io.Reader, ch chan<- Chunk) {
+	scanner := bufio.NewScanner(body)
+	scanner.Buffer(make([]byte, 0, 64*1024), 256*1024)
+
+	var fullText strings.Builder
+	var lastModel string
+
+	for scanner.Scan() {
+		line := scanner.Text()
+
+		// SSE 格式：data: {...}
+		if !strings.HasPrefix(line, "data: ") {
+			continue
+		}
+		data := strings.TrimPrefix(line, "data: ")
+		if data == "[DONE]" {
+			// 流结束，发送最终 chunk
+			ch <- Chunk{Delta: "", Done: true, Model: lastModel}
+			return
+		}
+
+		var delta streamDelta
+		if err := json.Unmarshal([]byte(data), &delta); err != nil {
+			o.logger.Warnw("llm: unmarshal delta failed", "error", err, "data", data)
+			continue
+		}
+
+		if delta.Model != "" {
+			lastModel = delta.Model
+		}
+
+		// 提取增量文本
+		if len(delta.Choices) > 0 {
+			content := delta.Choices[0].Delta.Content
+			if content != "" {
+				fullText.WriteString(content)
+				ch <- Chunk{Delta: content, Done: false, Model: lastModel}
+			}
+
+			// 某些模型在最后一个 choice 中携带 usage
+			if delta.Choices[0].FinishReason != nil && delta.Usage != nil {
+				ch <- Chunk{
+					Delta: "",
+					Done:  true,
+					Model: lastModel,
+					TokensUsed: &TokenUsage{
+						Prompt:     delta.Usage.PromptTokens,
+						Completion: delta.Usage.CompletionTokens,
+						Total:      delta.Usage.TotalTokens,
+					},
+				}
+				return
+			}
+		}
+	}
+
+	// scanner 结束但没收到 [DONE]
+	if err := scanner.Err(); err != nil {
+		o.logger.Warnw("llm: scan error", "error", err)
+	}
+	ch <- Chunk{Delta: "", Done: true, Model: lastModel}
+}
+
+// buildMessages 构建 OpenAI Chat API 的 messages 数组。
+func (o *OpenAIService) buildMessages(req Request) []chatMessage {
+	var messages []chatMessage
+
+	// System prompt
+	messages = append(messages, chatMessage{
+		Role:    "system",
+		Content: []contentPart{{Type: "text", Text: BuildSystemPrompt(req.Language, "")}},
+	})
+
+	// 历史消息
+	for _, msg := range req.History {
+		messages = append(messages, chatMessage{
+			Role:    msg.Role,
+			Content: []contentPart{{Type: "text", Text: msg.Content}},
+		})
+	}
+
+	// 当前用户消息（图像 + 文本）
+	var parts []contentPart
+	if len(req.Image) > 0 {
+		b64 := base64.StdEncoding.EncodeToString(req.Image)
+		parts = append(parts, contentPart{
+			Type:     "image_url",
+			ImageURL: &imageURL{URL: "data:image/jpeg;base64," + b64},
+		})
+	}
+	parts = append(parts, contentPart{Type: "text", Text: req.Text})
+	messages = append(messages, chatMessage{Role: "user", Content: parts})
+
+	return messages
+}

backend/internal/ai/llm/prompt.go

@@ -0,0 +1,25 @@
+package llm
+
+import "strings"
+
+// BuildSystemPrompt 根据语言和细节级别构建系统提示词。
+func BuildSystemPrompt(language, detailLevel string) string {
+	isChinese := strings.HasPrefix(language, "zh")
+
+	var prompt strings.Builder
+	if isChinese {
+		prompt.WriteString("你是一个视觉助手。用户通过摄像头看到一个场景，并用语音向你提问。请用简洁自然的中文回答。如果涉及视觉描述，先说\"我看到……\"。回答控制在3-5句话以内，除非用户要求详细说明。")
+	} else {
+		prompt.WriteString("You are a visual assistant. The user sees a scene through their camera and asks questions by voice. Answer concisely and naturally. If describing visual content, start with 'I see...'. Keep answers to 3-5 sentences unless the user asks for detail.")
+	}
+
+	if detailLevel == "high" {
+		if isChinese {
+			prompt.WriteString("请提供更详细的视觉描述，包括颜色、位置、数量等细节。")
+		} else {
+			prompt.WriteString(" Provide detailed visual descriptions including colors, positions, quantities, and other details.")
+		}
+	}
+
+	return prompt.String()
+}