WebRtc_QingGan/src/llm_stream.js

// 以流式方式请求LLM大模型接口，并打印流式返回内容

async function requestLLMStream({ apiKey, model, messages, onSegment }) {
  const response = await fetch('https://ark.cn-beijing.volces.com/api/v3/bots/chat/completions', {
    method: 'POST',
    headers: {
      'Authorization': `Bearer ${apiKey}`,
      'Content-Type': 'application/json',
      'Accept': 'text/event-stream',
      'Cache-Control': 'no-cache',
    },
    body: JSON.stringify({
      model,
      stream: true,
      stream_options: { include_usage: true },
      messages,
    }),
  });

  if (!response.ok) {
    throw new Error(`HTTP error! status: ${response.status}`);
  }

  const reader = response.body.getReader();
  const decoder = new TextDecoder('utf-8');
  let done = false;
  let buffer = '';
  let content = '';
  let pendingText = ''; // 待处理的文本片段

  // 分段分隔符
  const segmentDelimiters = /[，。：；！？,.:;!?]|\.{3,}|……|…/;

  while (!done) {
    const { value, done: doneReading } = await reader.read();
    done = doneReading;
    if (value) {
      const chunk = decoder.decode(value, { stream: true });
      buffer += chunk;
      
      // 处理SSE格式的数据
      const lines = buffer.split('\n');
      buffer = lines.pop(); // 最后一行可能是不完整的，留到下次
      
      for (const line of lines) {
        if (!line.trim()) continue;
        
        // 检查是否是SSE格式的数据行
        if (line.startsWith('data:')) {
          const jsonStr = line.substring(5).trim(); // 移除 'data:' 前缀
          
          if (jsonStr === '[DONE]') {
            console.log('LLM SSE流结束');
            // 处理最后的待处理文本（无论长度是否大于5个字）
            if (pendingText.trim() && onSegment) {
              console.log('处理最后的待处理文本:', pendingText.trim());
              await onSegment(pendingText.trim(), true);
            }
            continue;
          }
          
          try {
            const obj = JSON.parse(jsonStr);
            if (obj.choices && obj.choices[0] && obj.choices[0].delta && obj.choices[0].delta.content) {
              const deltaContent = obj.choices[0].delta.content;
              content += deltaContent;
              pendingText += deltaContent;
              console.log('LLM内容片段:', deltaContent);
              
              // 检查是否包含分段分隔符
              if (segmentDelimiters.test(pendingText)) {
                // 按分隔符分割文本
                const segments = pendingText.split(segmentDelimiters);
                
                // 重新组合处理：只处理足够长的完整段落
                let accumulatedText = '';
                let hasProcessed = false;
                
                for (let i = 0; i < segments.length - 1; i++) {
                  const segment = segments[i].trim();
                  if (segment) {
                    accumulatedText += segment;
                    // 找到分隔符
                    const delimiterMatch = pendingText.match(segmentDelimiters);
                    if (delimiterMatch) {
                      accumulatedText += delimiterMatch[0];
                    }
                    
                    // 如果累积文本长度大于5个字，处理它
                    if (accumulatedText.length > 6 && onSegment) {
                      console.log('检测到完整段落:', accumulatedText);
                      await onSegment(accumulatedText, false);
                      hasProcessed = true;
                      accumulatedText = ''; // 重置
                    }
                  }
                }
                
                // 更新pendingText
                if (hasProcessed) {
                  // 保留未处理的累积文本和最后一个不完整段落
                  pendingText = accumulatedText + (segments[segments.length - 1] || '');
                }
              }
            }
          } catch (e) {
            console.error('解析LLM SSE数据失败:', e, '原始数据:', jsonStr);
          }
        } else if (line.startsWith('event: ') || line.startsWith('id: ') || line.startsWith('retry: ')) {
          // 忽略SSE的其他字段
          continue;
        }
      }
    }
  }

  // 返回完整内容
  return content;
}

export { requestLLMStream };