@shun-js/aibaiban-server 0.8.4 → 0.8.6

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@shun-js/aibaiban-server",
-  "version": "0.8.4",
+  "version": "0.8.6",
   "description": "aibaiban.com server",
   "keywords": [
     "ai aibaiban"
@@ -44,5 +44,5 @@
     "access": "public",
     "registry": "https://registry.npmjs.org/"
   },
-  "gitHead": "52e8268b4ab1635eead49bf93fc2587c84e83fda"
+  "gitHead": "04eafc5cc7cc90e3fe81103daa88c168ff6f5037"
 }

package/server/controller/LLMController.js

@@ -5,18 +5,8 @@ const service = require('../service/LLMService.js');
  * controller
  */
 module.exports = (app) => {
-  // intent
-  app.post('/intent', (req, res) => {
-    service.intent(req, res);
-  });
-
-  // drawWithTools (使用结构化 JSON 输出)
-  app.post('/drawWithTools', (req, res) => {
-    service.drawWithTools(req, res);
-  });
-
-  // draw/agent (流式 Agent)
-  app.post('/draw/agent', (req, res) => {
+  // draw agent
+  app.post('/chat-streaming', (req, res) => {
     service.drawAgent(req, res);
   });
 };

package/server/service/LLMService.js

@@ -1,93 +1,7 @@
-// llm
-const { llmParseIntent } = require('../util/llm-intent.js');
-
-// structured json output
-const { generateFlowchartWithTools } = require('../util/llm-toolcall.js');
-
-// util
-const { chatFeishuMsg, chatResFeishuMsg, errorFeishuMsg } = require('../util/feishu.js');
-
 // llm agent
 const { callLLMForJSON, callLLM } = require('../util/llm-agent.js');
 const prompts = require('../util/prompt-agent.js');
 
-/**
- * intent
- * @param {*} req
- * @param {*} res
- * @returns
- */
-exports.intent = async (req, res) => {
-  const methodName = 'intent';
-
-  // check
-  if (!req.body.userPrompt) {
-    const msg = 'need userPrompt';
-    req.logger.error(methodName, msg);
-    res.jsonFail(msg);
-    return;
-  }
-
-  // const
-  const userPrompt = decodeURIComponent(req.body.userPrompt);
-  req.logger.info(methodName, 'userPrompt', userPrompt);
-  chatFeishuMsg(req);
-
-  // go
-  try {
-    const llmParseIntentRes = await llmParseIntent(userPrompt);
-    const llmParseIntentObj = JSON.parse(llmParseIntentRes);
-    req.logger.info(methodName, 'llmParseIntentObj', llmParseIntentObj);
-
-    // r
-    chatResFeishuMsg(req, JSON.stringify(llmParseIntentObj));
-    res.jsonSuccess('success', llmParseIntentObj);
-  } catch (error) {
-    const msg = 'parse intent error';
-    errorFeishuMsg(req, msg);
-    req.logger.error(methodName, msg, error);
-    res.jsonFail(msg);
-  }
-};
-
-/**
- * drawWithTools - 使用结构化 JSON 输出生成流程图
- * @param {*} req
- * @param {*} res
- * @returns
- */
-exports.drawWithTools = async (req, res) => {
-  const methodName = 'drawWithTools';
-
-  // check
-  if (!req.body.userPrompt) {
-    const msg = 'need userPrompt';
-    req.logger.error(methodName, msg);
-    res.jsonFail(msg);
-    return;
-  }
-
-  // const
-  const userPrompt = decodeURIComponent(req.body.userPrompt);
-  req.logger.info(methodName, 'userPrompt', userPrompt);
-  chatFeishuMsg(req);
-
-  // go
-  try {
-    const diagram = await generateFlowchartWithTools(userPrompt);
-    req.logger.info(methodName, 'diagram', diagram);
-
-    // r
-    chatResFeishuMsg(req, JSON.stringify(diagram));
-    res.jsonSuccess('success', diagram);
-  } catch (error) {
-    const msg = 'draw with tools error';
-    errorFeishuMsg(req, msg);
-    req.logger.error(methodName, msg, error);
-    res.jsonFail(msg);
-  }
-};
-
 /**
  * drawAgent - 流式 Agent 接口
  * 流程：router -> classify -> elaborate -> review -> generate
@@ -116,6 +30,8 @@ exports.drawAgent = async (req, res) => {
     const intentResult = await callLLMForJSON(prompts.ROUTER_PROMPT.replace('{input}', input), res, 'router');
     const intent = intentResult.intent;
     req.logger.info(methodName, 'intent', intent);
+    // 发送结果
+    res.streaming(`data: ${JSON.stringify({ step: 'router', intent })}\n\n`);
 
     // 非白板请求
     if (intent === 'irrelevant') {
@@ -132,6 +48,8 @@ exports.drawAgent = async (req, res) => {
     const classifyResult = await callLLMForJSON(prompts.CLASSIFY_PROMPT.replace('{input}', input), res, 'classify');
     const diagramType = classifyResult.diagramType;
     req.logger.info(methodName, 'diagramType', diagramType);
+    // 发送结果
+    res.streaming(`data: ${JSON.stringify({ step: 'classify', diagramType })}\n\n`);
 
     // 3. elaborate - 细化内容
     res.streaming(`data: ${JSON.stringify({ step: 'elaborate', status: 'start' })}\n\n`);
@@ -142,6 +60,8 @@ exports.drawAgent = async (req, res) => {
       'elaborate',
     );
     req.logger.info(methodName, 'elaboration', elaboration.slice(0, 100) + '...');
+    // 发送结果（不显示内容，只标记完成）
+    res.streaming(`data: ${JSON.stringify({ step: 'elaborate', done: true })}\n\n`);
 
     // 4. review - 质量检查
     res.streaming(`data: ${JSON.stringify({ step: 'review', status: 'start' })}\n\n`);
@@ -154,6 +74,8 @@ exports.drawAgent = async (req, res) => {
       'review',
     );
     req.logger.info(methodName, 'reviewResult', reviewResult);
+    // 发送结果
+    res.streaming(`data: ${JSON.stringify({ step: 'review', result: reviewResult.result })}\n\n`);
 
     // 信息不足，追问用户
     if (reviewResult.result === 'need_more_info') {

package/server/util/llm-agent.js

@@ -43,9 +43,9 @@ function extractJSON(text) {
 
 /**
  * 流式调用 LLM 并解析 JSON
- * 使用非流式 API，但过程流式输出给客户端
+ * 不发送中间 delta，只返回解析后的结果
  */
-exports.callLLMForJSON = async (prompt, res, step) => {
+exports.callLLMForJSON = async (prompt) => {
   // 非流式调用，等待完整响应
   const response = await llm.chat({
     model: llmConfig.modelName,
@@ -53,16 +53,15 @@ exports.callLLMForJSON = async (prompt, res, step) => {
   });
 
   const fullContent = response.content || '';
-  // 流式输出给客户端
-  res.streaming(`data: ${JSON.stringify({ step, delta: fullContent })}\n\n`);
-
+  // 不发送 delta，直接返回结果让调用方处理
   return extractJSON(fullContent);
 };
 
 /**
  * 流式调用 LLM（普通文本）
+ * 不发送中间 delta，只返回完整内容
  */
-exports.callLLM = async (prompt, res, step) => {
+exports.callLLM = async (prompt) => {
   let fullContent = '';
 
   await llm.chatWithStreaming(
@@ -70,8 +69,7 @@ exports.callLLM = async (prompt, res, step) => {
     {
       contentCallback: (chunk) => {
         fullContent += chunk;
-        // 直接流式返回给客户端
-        res.streaming(`data: ${JSON.stringify({ step, delta: chunk })}\n\n`);
+        // 不发送中间 delta，避免显示问题
       },
     },
   );