@templmf/temp-solf-lmf 0.0.43 → 0.0.45

package/ai-gateway/src/routes/chat.js

@@ -0,0 +1,657 @@
+/**
+ * POST /v1/chat/completions
+ *
+ * 完全兼容 OpenAI Chat Completions 格式的核心接口。
+ * 集成：Skill 自动加载 / 结构化响应 / Tool Use 适配 /
+ *       会话上下文管理 / 重试 / Streaming SSE
+ */
+
+import { Router } from "express";
+import OpenAI from "openai";
+import { v4 as uuidv4 } from "uuid";
+
+import {
+  detectSkills,
+  detectSkillsByKeyword,
+  detectSkillsWithAI,
+  buildSkillPrompt
+} from "../skills/registry.js";
+import {
+  getModelCapabilities,
+  adaptTools,
+  stripThinkingBlock,
+  parseToolFallbackResponse
+} from "../utils/toolAdapter.js";
+import {
+  buildResponseFormatPrompt,
+  parseStructuredResponse,
+  wrapStructuredChoice
+} from "../utils/structuredResponse.js";
+import {
+  createSession,
+  getSession,
+  getSessionMessages,
+  appendMessages,
+  updateSessionMetadata,
+  compressIfNeeded
+} from "../utils/sessionManager.js";
+import {
+  withRetry,
+  buildJsonCorrectionMessages
+} from "../utils/retry.js";
+import { logger } from "../utils/logger.js";
+
+export const chatRouter = Router();
+
+// ─────────────────────────────────────────────────────
+// 上游模型客户端（内网 Qwen 服务）
+// ─────────────────────────────────────────────────────
+const upstreamClient = new OpenAI({
+  apiKey: process.env.UPSTREAM_API_KEY || "none",
+  baseURL: process.env.UPSTREAM_BASE_URL || "http://localhost:8000/v1",
+  timeout: parseInt(process.env.UPSTREAM_TIMEOUT_MS || "120000"),
+  maxRetries: 0
+});
+
+// ─────────────────────────────────────────────────────
+// POST /v1/chat/completions
+// ─────────────────────────────────────────────────────
+chatRouter.post("/", async (req, res) => {
+  const requestId = uuidv4();
+  const startTime = Date.now();
+
+  // ── 1. 解析请求体 ─────────────────────────────────
+  const {
+    model,
+    messages: clientMessages,
+    system,
+    tools,
+    tool_choice,
+    stream = false,
+    max_tokens,
+    temperature,
+    top_p,
+    session_id,
+    response_format,
+    x_skill_override,
+    x_keep_thinking = false
+  } = req.body;
+
+  if (!clientMessages || !Array.isArray(clientMessages) || clientMessages.length === 0) {
+    return res.status(400).json({
+      error: {
+        message: "messages is required and must be a non-empty array",
+        type: "invalid_request_error"
+      }
+    });
+  }
+
+  const resolvedModel = model || process.env.DEFAULT_MODEL || "qwen2.5-72b-instruct";
+  const caps = getModelCapabilities(resolvedModel);
+
+  // ── 2. 会话管理 ───────────────────────────────────
+  let resolvedSessionId = session_id;
+  let historyMessages = [];
+
+  if (resolvedSessionId) {
+    const session = getSession(resolvedSessionId);
+    if (session) {
+      historyMessages = getSessionMessages(resolvedSessionId);
+    } else {
+      resolvedSessionId = createSession(req.clientId);
+      logger.info("Session expired, created new", { requestId, newSessionId: resolvedSessionId });
+    }
+  } else {
+    resolvedSessionId = createSession(req.clientId);
+  }
+
+  const allMessages = [...historyMessages, ...clientMessages];
+  const { messages: finalMessages, compressed } = compressIfNeeded(
+    allMessages,
+    parseInt(process.env.MAX_CONTEXT_TOKENS || "6000")
+  );
+
+  // ── 3. Skill 检测 ─────────────────────────────────
+  // 统一使用 skillDetectMethod 这一个变量名贯穿全文
+  let skillNames = [];
+  let skillDetectMethod = "none";
+
+  if (x_skill_override) {
+    skillNames = Array.isArray(x_skill_override) ? x_skill_override : [x_skill_override];
+    skillDetectMethod = "override";
+  } else if (process.env.USE_AI_SKILL_DETECTION === "true") {
+    // AI 语义检测
+    const aiResult = await detectSkillsWithAI(finalMessages, upstreamClient, resolvedModel);
+    skillNames = aiResult.names;
+    skillDetectMethod = skillNames.length > 0 ? "ai" : "none";
+  } else {
+    // 关键词匹配（默认，支持组合规则）
+    skillNames = detectSkills(finalMessages);
+    skillDetectMethod = skillNames.length > 0 ? "keyword" : "none";
+  }
+
+  const skillPrompt = buildSkillPrompt(skillNames);
+
+  // ── 4. 结构化响应格式 prompt ──────────────────────
+  const formatPrompt = buildResponseFormatPrompt(response_format);
+
+  // ── 5. 组装 system prompt ─────────────────────────
+  const baseSystem = system || "你是一个智能助手，擅长理解用户需求并提供精准、实用的回答。";
+  const finalSystem = `${baseSystem}${skillPrompt}${formatPrompt}`;
+
+  // ── 6. Tool Use 适配 ──────────────────────────────
+  const { tools: adaptedTools, toolFallbackPrompt } = adaptTools(tools, resolvedModel);
+  const systemWithTools = toolFallbackPrompt
+    ? finalSystem + toolFallbackPrompt
+    : finalSystem;
+
+  // ── 7. 上游请求参数 ───────────────────────────────
+  const upstreamPayload = {
+    model: resolvedModel,
+    messages: [
+      { role: "system", content: systemWithTools },
+      ...finalMessages
+    ],
+    stream,
+    ...(adaptedTools?.length && { tools: adaptedTools }),
+    ...(tool_choice && adaptedTools?.length && { tool_choice }),
+    ...(max_tokens && { max_tokens }),
+    ...(temperature !== undefined && { temperature }),
+    ...(top_p !== undefined && { top_p }),
+    ...(caps.requiresThinkingBlock && {
+      extra_body: { enable_thinking: true }
+    })
+  };
+
+  logger.info("Chat request", {
+    requestId,
+    clientId: req.clientId,
+    sessionId: resolvedSessionId,
+    model: resolvedModel,
+    historyCount: historyMessages.length,
+    skillsDetected: skillNames,
+    skillDetectMethod,
+    stream,
+    responseFormat: response_format?.type || response_format || null,
+    contextCompressed: compressed
+  });
+
+  // ── 8. 路由 ───────────────────────────────────────
+  if (stream) {
+    return handleStream(res, upstreamClient, upstreamPayload, {
+      requestId, resolvedModel, resolvedSessionId,
+      skillNames, skillDetectMethod, clientMessages, x_keep_thinking, caps
+    });
+  }
+
+  return handleNonStream(res, upstreamClient, upstreamPayload, {
+    requestId, resolvedModel, resolvedSessionId,
+    skillNames, skillDetectMethod, clientMessages, finalMessages,
+    response_format, x_keep_thinking, caps, startTime
+  });
+});
+
+// ─────────────────────────────────────────────────────
+// 非流式响应处理
+// ─────────────────────────────────────────────────────
+async function handleNonStream(res, client, payload, ctx) {
+  const {
+    requestId, resolvedModel, resolvedSessionId,
+    skillNames, skillDetectMethod, clientMessages, finalMessages,
+    response_format, x_keep_thinking, caps, startTime
+  } = ctx;
+
+  const responseFormatType = response_format?.type || response_format;
+  const isJsonFormat = ["json_object", "json_schema", "ticket_form"].includes(responseFormatType);
+
+  try {
+    let completion = await withRetry(
+      () => client.chat.completions.create(payload),
+      { requestId, label: "upstream_call", maxRetries: 2 }
+    );
+
+    // 过滤 thinking block
+    let processedChoices = completion.choices.map(choice => {
+      if (choice.message?.content) {
+        choice.message.content = stripThinkingBlock(
+          choice.message.content, x_keep_thinking
+        );
+      }
+      return choice;
+    });
+
+    // Tool fallback 解析
+    if (!payload.tools && systemWithToolsCheck(payload)) {
+      processedChoices = processedChoices.map(choice => {
+        const fallback = parseToolFallbackResponse(choice.message?.content || "");
+        return fallback ? { ...choice, message: fallback } : choice;
+      });
+    }
+
+    // 结构化输出解析，失败时追加纠正重试
+    if (response_format && isJsonFormat) {
+      const firstParseResult = parseStructuredResponse(
+        processedChoices[0]?.message?.content || "",
+        response_format
+      );
+
+      if (!firstParseResult.success) {
+        logger.warn("Structured parse failed, retrying with correction", { requestId });
+
+        const correctedMessages = buildJsonCorrectionMessages(
+          finalMessages,
+          processedChoices[0]?.message?.content || "",
+          response_format.schema
+        );
+
+        const correctionPayload = {
+          ...payload,
+          messages: [payload.messages[0], ...correctedMessages]
+        };
+
+        const corrected = await withRetry(
+          () => client.chat.completions.create(correctionPayload),
+          { requestId, label: "json_correction", maxRetries: 1 }
+        );
+
+        completion = corrected;
+        processedChoices = corrected.choices.map(choice => {
+          if (choice.message?.content) {
+            choice.message.content = stripThinkingBlock(
+              choice.message.content, x_keep_thinking
+            );
+          }
+          return choice;
+        });
+      }
+
+      processedChoices = processedChoices.map(choice => {
+        const result = parseStructuredResponse(
+          choice.message?.content || "", response_format
+        );
+        return wrapStructuredChoice(choice, result, response_format);
+      });
+
+      if (responseFormatType === "ticket_form") {
+        const parsed = processedChoices[0]?.message?.parsed;
+        if (parsed?.form_draft) {
+          updateSessionMetadata(resolvedSessionId, { ticket_draft: parsed.form_draft });
+        }
+      }
+    }
+
+    // 保存到 session
+    appendMessages(resolvedSessionId, [
+      ...clientMessages,
+      {
+        role: "assistant",
+        content: processedChoices[0]?.message?.content || ""
+      }
+    ]);
+
+    const matchedMcps = extractMcpMatches(processedChoices[0]?.message?.tool_calls);
+
+    const response = {
+      ...completion,
+      choices: processedChoices,
+      _gateway: {
+        request_id: requestId,
+        session_id: resolvedSessionId,
+        matched: buildMatchedField(skillNames, skillDetectMethod, matchedMcps),
+        skills_loaded: skillNames,
+        model_caps: {
+          toolSupport: caps.toolSupport,
+          requiresThinkingBlock: caps.requiresThinkingBlock
+        },
+        latency_ms: Date.now() - startTime
+      }
+    };
+
+    logger.info("Chat completed", {
+      requestId,
+      sessionId: resolvedSessionId,
+      model: resolvedModel,
+      skillsLoaded: skillNames,
+      latencyMs: Date.now() - startTime,
+      promptTokens: completion.usage?.prompt_tokens,
+      completionTokens: completion.usage?.completion_tokens
+    });
+
+    res.json(response);
+
+  } catch (err) {
+    handleUpstreamError(err, res, requestId);
+  }
+}
+
+// ─────────────────────────────────────────────────────
+// 流式响应处理（SSE）
+// ─────────────────────────────────────────────────────
+async function handleStream(res, client, payload, ctx) {
+  const {
+    requestId, resolvedModel, resolvedSessionId,
+    skillNames, skillDetectMethod, clientMessages, x_keep_thinking
+  } = ctx;
+
+  res.setHeader("Content-Type", "text/event-stream");
+  res.setHeader("Cache-Control", "no-cache");
+  res.setHeader("Connection", "keep-alive");
+  res.setHeader("X-Gateway-Request-Id", requestId);
+  res.setHeader("X-Gateway-Session-Id", resolvedSessionId);
+  res.setHeader("X-Gateway-Skills", skillNames.join(","));
+  res.setHeader("X-Gateway-Skill-Detect", skillDetectMethod);
+  res.flushHeaders();
+
+  // 首个事件：推送 session 信息
+  res.write(`data: ${JSON.stringify({
+    id: requestId,
+    object: "chat.completion.chunk",
+    type: "session",
+    session_id: resolvedSessionId,
+    matched: buildMatchedField(skillNames, skillDetectMethod, []),
+    skills_loaded: skillNames
+  })}\n\n`);
+
+  let fullContent = "";
+  let insideThinkBlock = false;
+  const toolCallsAcc = {};
+
+  try {
+    const streamResponse = await withRetry(
+      () => client.chat.completions.create(payload),
+      { requestId, label: "stream_request", maxRetries: 1 }
+    );
+
+    for await (const chunk of streamResponse) {
+      const delta = chunk.choices?.[0]?.delta || {};
+
+      // 累积 tool_calls 分片
+      if (delta.tool_calls) {
+        for (const tc of delta.tool_calls) {
+          const idx = tc.index ?? 0;
+          if (!toolCallsAcc[idx]) {
+            toolCallsAcc[idx] = {
+              id:       tc.id || "",
+              type:     "function",
+              function: { name: "", arguments: "" }
+            };
+          }
+          if (tc.id)                  toolCallsAcc[idx].id                   = tc.id;
+          if (tc.function?.name)      toolCallsAcc[idx].function.name       += tc.function.name;
+          if (tc.function?.arguments) toolCallsAcc[idx].function.arguments  += tc.function.arguments;
+        }
+      }
+
+      // 文本 delta 过滤与透传
+      const textDelta = delta.content || "";
+      if (textDelta && !x_keep_thinking) {
+        if (textDelta.includes("<think>")) insideThinkBlock = true;
+        if (insideThinkBlock) {
+          if (textDelta.includes("</think>")) insideThinkBlock = false;
+          continue;
+        }
+      }
+
+      fullContent += textDelta;
+      res.write(`data: ${JSON.stringify(chunk)}\n\n`);
+    }
+
+    const cleanContent  = stripThinkingBlock(fullContent, x_keep_thinking);
+    const toolCallsList = Object.values(toolCallsAcc).filter(tc => tc.function?.name);
+
+    // tool_calls 由客户端负责执行并管理消息历史。
+    // 网关 session 只存纯文本回复，不存 tool_calls，
+    // 避免下一轮 session 历史里出现没有对应 tool 回复的悬空 tool_calls。
+    if (toolCallsList.length === 0) {
+      appendMessages(resolvedSessionId, [
+        ...clientMessages,
+        { role: "assistant", content: cleanContent || "" }
+      ]);
+    }
+    // 有 tool_calls 时不写 session，客户端自行维护完整的工具调用轮次
+
+    res.write("data: [DONE]\n\n");
+    res.end();
+
+    logger.info("Stream completed", {
+      requestId,
+      sessionId:     resolvedSessionId,
+      model:         resolvedModel,
+      skillsLoaded:  skillNames,
+      toolCallCount: toolCallsList.length
+    });
+
+  } catch (err) {
+    handleUpstreamStreamError(err, res, requestId);
+  }
+}
+
+// ─────────────────────────────────────────────────────
+// 辅助：判断是否启用了 tool fallback prompt
+// ─────────────────────────────────────────────────────
+function systemWithToolsCheck(payload) {
+  const sys = payload.messages?.[0]?.content || "";
+  return sys.includes("tool_call") && sys.includes("工具名");
+}
+
+// ─────────────────────────────────────────────────────
+// 上游错误翻译
+// 把上游原始错误转成对用户友好的结构化信息
+// ─────────────────────────────────────────────────────
+function translateUpstreamError(err) {
+  const raw     = err.message || "";
+  const status  = err.status  || err.statusCode || 500;
+  const code    = err.code    || null;
+
+  // ── 配额 / 免费层耗尽 ────────────────────────────
+  if (
+    status === 403 ||
+    raw.includes("free tier") ||
+    raw.includes("has been exhausted") ||
+    raw.includes("quota") ||
+    raw.includes("insufficient_quota") ||
+    code === "insufficient_quota"
+  ) {
+    return {
+      status:  403,
+      message: "上游模型配额已耗尽（免费额度用完）。请在模型管理控制台关闭「仅免费层」模式，或充值后重试。",
+      type:    "quota_exceeded",
+      code:    "upstream_quota_exhausted",
+      hint:    raw   // 保留原始信息供排查
+    };
+  }
+
+  // ── 认证失败 ─────────────────────────────────────
+  if (
+    status === 401 ||
+    raw.includes("Incorrect API key") ||
+    raw.includes("invalid_api_key") ||
+    code === "invalid_api_key"
+  ) {
+    return {
+      status:  401,
+      message: "上游 API Key 无效或已过期，请检查 UPSTREAM_API_KEY 配置。",
+      type:    "auth_error",
+      code:    "upstream_auth_failed",
+      hint:    raw
+    };
+  }
+
+  // ── 模型不存在 ────────────────────────────────────
+  if (
+    status === 404 ||
+    raw.includes("model") && raw.includes("not found") ||
+    raw.includes("does not exist") ||
+    code === "model_not_found"
+  ) {
+    return {
+      status:  404,
+      message: `上游模型不存在或未部署，请检查 DEFAULT_MODEL / model 参数配置。原始信息：${raw}`,
+      type:    "model_not_found",
+      code:    "upstream_model_not_found",
+      hint:    raw
+    };
+  }
+
+  // ── 限流 ─────────────────────────────────────────
+  if (
+    status === 429 ||
+    raw.includes("rate limit") ||
+    raw.includes("too many requests") ||
+    code === "rate_limit_exceeded"
+  ) {
+    return {
+      status:  429,
+      message: "上游服务限流，请求过于频繁，请稍后重试。",
+      type:    "rate_limit_error",
+      code:    "upstream_rate_limit",
+      hint:    raw
+    };
+  }
+
+  // ── 上游超时 ─────────────────────────────────────
+  if (
+    status === 408 || status === 504 ||
+    raw.includes("timeout") ||
+    raw.includes("timed out") ||
+    err.code === "ECONNABORTED"
+  ) {
+    return {
+      status:  504,
+      message: "上游模型响应超时，请稍后重试或减少输入长度。",
+      type:    "timeout_error",
+      code:    "upstream_timeout",
+      hint:    raw
+    };
+  }
+
+  // ── 上游服务不可达 ────────────────────────────────
+  if (
+    status >= 500 && status < 600 ||
+    err.code === "ECONNREFUSED" ||
+    err.code === "ENOTFOUND"
+  ) {
+    return {
+      status:  502,
+      message: `上游服务暂时不可用（${status}），请检查 UPSTREAM_BASE_URL 或稍后重试。`,
+      type:    "upstream_error",
+      code:    "upstream_unavailable",
+      hint:    raw
+    };
+  }
+
+  // ── 其他未知错误（原样返回，但保证格式统一）─────
+  return {
+    status:  status < 100 ? 500 : status,
+    message: raw || "未知错误，请稍后重试。",
+    type:    err.type || "api_error",
+    code:    code,
+    hint:    null
+  };
+}
+
+// ─────────────────────────────────────────────────────
+// 上游错误统一处理（非流式）
+// ─────────────────────────────────────────────────────
+function handleUpstreamError(err, res, requestId) {
+  const translated = translateUpstreamError(err);
+
+  logger.error("Upstream error", {
+    requestId,
+    status:   translated.status,
+    type:     translated.type,
+    code:     translated.code,
+    rawError: translated.hint || err.message
+  });
+
+  res.status(translated.status).json({
+    error: {
+      message:    translated.message,
+      type:       translated.type,
+      code:       translated.code,
+      request_id: requestId,
+      ...(translated.hint && process.env.NODE_ENV !== "production"
+        ? { upstream_detail: translated.hint }
+        : {})
+    }
+  });
+}
+
+// ─────────────────────────────────────────────────────
+// 上游错误统一处理（流式 SSE）
+// SSE 已经发了 headers，只能通过 data 帧传递错误
+// ─────────────────────────────────────────────────────
+function handleUpstreamStreamError(err, res, requestId) {
+  const translated = translateUpstreamError(err);
+
+  logger.error("Stream upstream error", {
+    requestId,
+    status:   translated.status,
+    type:     translated.type,
+    code:     translated.code,
+    rawError: translated.hint || err.message
+  });
+
+  res.write(`data: ${JSON.stringify({
+    error: {
+      message:    translated.message,
+      type:       translated.type,
+      code:       translated.code,
+      request_id: requestId,
+      ...(translated.hint && process.env.NODE_ENV !== "production"
+        ? { upstream_detail: translated.hint }
+        : {})
+    }
+  })}\n\n`);
+  res.end();
+}
+
+// ─────────────────────────────────────────────────────
+// 从 tool_calls 里提取命中的 MCP server 信息
+// mcp__<server-name>__<tool-name> 格式
+// ─────────────────────────────────────────────────────
+function extractMcpMatches(toolCalls) {
+  if (!toolCalls || toolCalls.length === 0) return [];
+
+  const mcpMap = {};
+  for (const tc of toolCalls) {
+    const name  = tc?.function?.name || tc?.name || "";
+    const match = name.match(/^mcp__([^_]+(?:_[^_]+)*)__(.+)$/);
+    if (match) {
+      const serverName = match[1];
+      const toolName   = match[2];
+      if (!mcpMap[serverName]) {
+        mcpMap[serverName] = { server: serverName, tools_called: [] };
+      }
+      mcpMap[serverName].tools_called.push(toolName);
+    }
+  }
+
+  return Object.values(mcpMap);
+}
+
+// ─────────────────────────────────────────────────────
+// 构建 matched 字段
+// 只在有命中时填充，未命中时返回 null，流程照常继续
+// ─────────────────────────────────────────────────────
+function buildMatchedField(skillNames, detectMethod, matchedMcps) {
+  const hasSkills = skillNames && skillNames.length > 0;
+  const hasMcps   = matchedMcps && matchedMcps.length > 0;
+
+  if (!hasSkills && !hasMcps) return null;
+
+  return {
+    ...(hasSkills && {
+      skills: skillNames.map(name => ({
+        name,
+        detection_method: detectMethod  // "keyword" | "ai" | "override"
+      }))
+    }),
+    ...(hasMcps && {
+      mcps: matchedMcps.map(m => ({
+        server:       m.server,
+        tools_called: m.tools_called
+      }))
+    })
+  };
+}