Upload 64 files

ai-context.js

@@ -1,8 +1,5 @@
 
-const { 
-    User, Student, Score, AttendanceModel, ClassModel, 
-    LeaveRequestModel, TodoModel, School, Course 
-} = require('./models');
+const { User, Student, School } = require('./models');
 
 /**
  * 格式化当前日期
@@ -14,259 +11,50 @@ const getCurrentDateInfo = () => {
 };
 
 /**
- * 构建学生画像上下文 (学生视角)
- */
-async function buildStudentContext(username, schoolId) {
-    const student = await Student.findOne({ 
-        $or: [{ studentNo: username }, { name: username }],
-        schoolId 
-    });
-
-    if (!student) return "无法找到该学生的详细档案。";
-
-    // 1. 获取近期成绩 (最近10条，让AI掌握更多趋势)
-    const recentScores = await Score.find({ 
-        studentNo: student.studentNo, 
-        schoolId 
-    }).sort({ _id: -1 }).limit(10);
-
-    // 2. 获取考勤概况
-    const attendanceStats = await AttendanceModel.aggregate([
-        { $match: { studentId: student._id.toString() } },
-        { $group: { _id: "$status", count: { $sum: 1 } } }
-    ]);
-    const absentCount = attendanceStats.find(a => a._id === 'Absent')?.count || 0;
-    const leaveCount = attendanceStats.find(a => a._id === 'Leave')?.count || 0;
-
-    // 3. 获取待办事项
-    const user = await User.findOne({ username, schoolId });
-    const todos = user ? await TodoModel.find({ userId: user._id, isCompleted: false }).limit(5) : [];
-
-    let prompt = `
-### 当前用户身份：学生 (个人视图)
-- **姓名**: ${student.name}
-- **班级**: ${student.className}
-- **学号**: ${student.studentNo}
-- **积分(小红花)**: ${student.flowerBalance} 🌺
-
-### 个人学习数据
-`;
-
-    if (recentScores.length > 0) {
-        prompt += `- **近期成绩历史**: ${recentScores.map(s => `${s.courseName}: ${s.score} (${s.type || '考试'})`).join('; ')}\n`;
-        // 计算简单平均分
-        const avg = (recentScores.reduce((acc, s) => acc + s.score, 0) / recentScores.length).toFixed(1);
-        prompt += `- **近期平均分**: ${avg}\n`;
-    } else {
-        prompt += `- **近期成绩**: 暂无记录\n`;
-    }
-
-    if (absentCount > 0 || leaveCount > 0) {
-        prompt += `- **考勤异常**: 本学期缺勤 ${absentCount} 次，请假 ${leaveCount} 次。\n`;
-    } else {
-        prompt += `- **考勤状况**: 全勤，表现极佳。\n`;
-    }
-
-    if (todos.length > 0) {
-        prompt += `- **未完成待办**: ${todos.map(t => t.content).join('; ')}\n`;
-    }
-
-    return prompt;
-}
-
-/**
- * 构建教师画像上下文 (增强版 - 智能区分班主任与科任视角)
- */
-async function buildTeacherContext(username, schoolId) {
-    const user = await User.findOne({ username, schoolId });
-    if (!user) return "无法找到该教师档案。";
-
-    // 1. 确定老师的身份范围
-    const homeroomClassName = user.homeroomClass; // 班主任班级
-    
-    // 查找该老师任教的所有课程 (找出任教的其他班级)
-    const teachingCourses = await Course.find({ 
-        $or: [{ teacherId: user._id }, { teacherName: user.trueName || user.username }],
-        schoolId 
-    });
-    
-    // 构建任教班级 -> 科目列表的映射 (e.g., "三年级(2)班": ["数学", "科学"])
-    const teachingMap = {};
-    teachingCourses.forEach(c => {
-        if (!teachingMap[c.className]) teachingMap[c.className] = new Set();
-        teachingMap[c.className].add(c.courseName);
-    });
-
-    // 合并所有相关班级 (班主任班级 + 任课班级)
-    const allClasses = new Set(Object.keys(teachingMap));
-    if (homeroomClassName) allClasses.add(homeroomClassName);
-
-    if (allClasses.size === 0) {
-        return `### 当前用户身份：教师\n- **姓名**: ${user.trueName || username}\n- **状态**: 暂未绑定任何班级或课程数据。`;
-    }
-
-    let prompt = `
-### 当前用户身份：教师
-- **姓名**: ${user.trueName || username}
-- **负责班级**: ${Array.from(allClasses).join(', ')}
-`;
-
-    // 2. 遍历所有相关班级，构建详细数据
-    for (const className of allClasses) {
-        const isHomeroom = className === homeroomClassName;
-        const subjectsTaught = teachingMap[className] ? Array.from(teachingMap[className]) : [];
-        
-        prompt += `\n#### 🏫 班级: ${className} (${isHomeroom ? '我是班主任' : '我是任课老师'})\n`;
-        if (!isHomeroom) {
-            prompt += `(非班主任视角：仅展示我任教的科目 [${subjectsTaught.join(', ')}] 的数据)\n`;
-        }
-
-        // 2.1 获取该班学生
-        const students = await Student.find({ className, schoolId });
-        if (students.length === 0) {
-            prompt += `- 暂无学生数据\n`;
-            continue;
-        }
-        
-        const studentNos = students.map(s => s.studentNo);
-        const studentIds = students.map(s => s._id.toString());
-
-        // 2.2 获取考勤 (全班)
-        const attendanceRaw = await AttendanceModel.aggregate([
-            { $match: { studentId: { $in: studentIds }, status: { $in: ['Absent', 'Leave'] } } },
-            { $group: { _id: "$studentId", absent: { $sum: { $cond: [{ $eq: ["$status", "Absent"] }, 1, 0] } }, leave: { $sum: { $cond: [{ $eq: ["$status", "Leave"] }, 1, 0] } } } }
-        ]);
-        const attendanceMap = {};
-        attendanceRaw.forEach(a => attendanceMap[a._id] = a);
-
-        // 2.3 获取成绩 (按需获取)
-        // 查询该班级学生的所有成绩
-        // 为了性能，还是查出来再内存过滤，比多次DB查询快
-        const allScores = await Score.find({ 
-            schoolId, 
-            studentNo: { $in: studentNos }
-        }).sort({ _id: -1 }); // 最新的在前
-
-        // 构建每个学生的成绩摘要
-        const studentDetails = students.map(s => {
-            const att = attendanceMap[s._id.toString()] || { absent: 0, leave: 0 };
-            
-            // 筛选该学生的成绩
-            let myScores = allScores.filter(sc => sc.studentNo === s.studentNo);
-
-            // 【关键逻辑】过滤显示哪些科目
-            if (!isHomeroom) {
-                // 如果不是班主任，只保留我教的科目的成绩
-                myScores = myScores.filter(sc => subjectsTaught.includes(sc.courseName));
-            }
-
-            // 【聚合逻辑】每个科目只取最近一次成绩 (去重)
-            const latestSubjectScores = {};
-            myScores.forEach(sc => {
-                if (!latestSubjectScores[sc.courseName]) {
-                    latestSubjectScores[sc.courseName] = sc;
-                }
-            });
-
-            const finalScores = Object.values(latestSubjectScores);
-            
-            // 格式化成绩字符串
-            let scoreStr = "";
-            if (finalScores.length > 0) {
-                scoreStr = finalScores.map(sc => `${sc.courseName}:${sc.score}`).join(', ');
-            } else {
-                scoreStr = "无相关成绩";
-            }
-
-            // 标记异常 (缺勤多 或 有不及格)
-            const hasIssue = att.absent > 0 || finalScores.some(sc => sc.score < 60);
-            const flag = hasIssue ? "⚠️" : "";
-
-            return `- ${flag} **${s.name}**: 考勤[缺${att.absent}/假${att.leave}], 小红花:${s.flowerBalance}, 最新成绩:[${scoreStr}]`;
-        });
-
-        // 将学生列表加入 Prompt (限制长度，如果班级人太多，可能需要截断，但Gemini窗口大，通常没事)
-        prompt += studentDetails.join('\n') + '\n';
-
-        // 2.4 如果是班主任，额外显示待办
-        if (isHomeroom) {
-            const pendingLeaves = await LeaveRequestModel.countDocuments({ className, schoolId, status: 'Pending' });
-            if (pendingLeaves > 0) {
-                prompt += `> 🔴 班务提醒: 有 ${pendingLeaves} 条请假申请待审批。\n`;
-            }
-        }
-    }
-
-    return prompt;
-}
-
-/**
- * 构建管理员/校长画像上下文
- */
-async function buildAdminContext(role, schoolId) {
-    let prompt = `### 当前用户身份：${role === 'PRINCIPAL' ? '校长' : '超级管理员'}\n`;
-    
-    if (role === 'PRINCIPAL' && schoolId) {
-        const school = await School.findById(schoolId);
-        const totalStudents = await Student.countDocuments({ schoolId });
-        const totalTeachers = await User.countDocuments({ schoolId, role: 'TEACHER' });
-        
-        // 今日缺勤详细名单
-        const today = new Date().toISOString().split('T')[0];
-        const absences = await AttendanceModel.find({ schoolId, date: today, status: { $in: ['Absent', 'Leave'] } });
-        const absentNames = absences.map(a => `${a.studentName}(${a.className})`).join(', ');
-
-        // 全校均分
-        const recentScores = await Score.find({ schoolId }).sort({_id:-1}).limit(100);
-        let avgScore = 0;
-        if (recentScores.length) avgScore = (recentScores.reduce((a,b)=>a+b.score,0)/recentScores.length).toFixed(1);
-
-        prompt += `- **学校**: ${school ? school.name : '未知'}\n`;
-        prompt += `- **宏观数据**: 教师 ${totalTeachers} 人，学生 ${totalStudents} 人，近期全校抽样平均分 ${avgScore}。\n`;
-        prompt += `- **今日出勤**: 缺勤/请假 ${absences.length} 人。名单: ${absentNames || '无'}。\n`;
-    }
-
-    return prompt;
-}
-
-/**
- * 主入口：构建用户上下文 Prompt
- * @param {string} username - 请求头中的用户名
- * @param {string} role - 请求头中的角色
- * @param {string} schoolId - 请求头中的学校ID
+ * 构建用户上下文 - Agentic版 (精简)
+ * 既然 AI 现在有了 query_database 工具，我们不需要把所有数据都塞进 System Prompt。
+ * 我们只需要告诉它：“你是谁”，“用户是谁”，以及“你有查库的能力”。
  */
 async function buildUserContext(username, role, schoolId) {
     try {
         const dateStr = getCurrentDateInfo();
-        let roleContext = "";
+        let userProfile = "";
 
+        // 基础用户信息
         if (role === 'STUDENT') {
-            roleContext = await buildStudentContext(username, schoolId);
+            const student = await Student.findOne({ 
+                $or: [{ studentNo: username }, { name: username }],
+                schoolId 
+            });
+            if (student) {
+                userProfile = `用户是学生：${student.name} (班级: ${student.className}, 学号: ${student.studentNo})`;
+            }
         } else if (role === 'TEACHER') {
-            roleContext = await buildTeacherContext(username, schoolId);
-        } else if (role === 'ADMIN' || role === 'PRINCIPAL') {
-            roleContext = await buildAdminContext(role, schoolId);
+            const user = await User.findOne({ username, schoolId });
+            if (user) {
+                userProfile = `用户是教师：${user.trueName || username} (任教: ${user.teachingSubject || '无'}, 班主任: ${user.homeroomClass || '否'})`;
+            }
+        } else {
+            userProfile = `用户是管理员/校长。`;
         }
 
-        // 组装最终 System Instruction 片段
         return `
 ---
-【上下文注入信息 (Context Injection) - 绝密】
-当前系统时间: ${dateStr}
-以下是当前用户的核心数据和其管辖范围内的详细档案。
-${roleContext}
-
-【AI 行为准则】
-1. 你拥有上述所有数据的“上帝视角”。
-2. **班主任视角**: 当用户是班主任时，你通过上下文已知晓全班所有科目的成绩。如果问“王五偏科吗”，请对比他的各科成绩作答。
-3. **任课老师视角**: 当用户非班主任时，你只能看到他所教科目的成绩。如果问“李华其他课怎么样”，请诚实回答“我只能看到您任教科目的数据，无法评价其他科目”。
-4. 回答要具体。不要说“他成绩一般”，要说“他最近数学考了60分，英语考了85分”。
-5. 数据格式说明: [科目:分数] 代表该科目最近一次录入的成绩。
+【系统信息】
+当前时间: ${dateStr}
+${userProfile}
+
+【能力说明】
+1. 你是一个拥有“本地数据库查询权限”的智能助教。
+2. 工具执行环境：**工具是在用户的本地服务器上执行的**。你可以通过 Function Calling 获取内网数据，**不需要**公网访问权限。
+3. 如果用户问“张三考了多少分”或“我们班谁考勤不好”，请**务必**大胆调用 \`query_database\` 工具。
+4. 不要回答“我无法访问数据库”，因为工具会帮你完成访问并将结果传回给你。
+5. 如果查询结果返回 JSON 数据，请将其整理为通俗易懂的自然语言回答用户。
 ---
 `;
     } catch (e) {
         console.error("Context build failed:", e);
-        return ""; // 失败时降级为空，不影响主流程
+        return "";
     }
 }
 

ai-routes.js

@@ -4,8 +4,8 @@ const router = express.Router();
 const OpenAI = require('openai');
 const { ConfigModel, User, AIUsageModel, ChatHistoryModel } = require('./models');
 const { buildUserContext } = require('./ai-context');
+const { mongoTools, getOpenAITools, executeMongoTool } = require('./ai-tools');
 
-// ... (Key Management, Usage Tracking, Helpers remain same)
 // Fetch keys from DB + merge with ENV variables
 async function getKeyPool(type) {
     const config = await ConfigModel.findOne({ key: 'main' });
@@ -26,266 +26,6 @@ async function recordUsage(model, provider) {
     } catch (e) { console.error("Failed to record AI usage stats:", e); }
 }
 
-const wait = (ms) => new Promise(resolve => setTimeout(resolve, ms));
-async function callAIWithRetry(aiModelCall, retries = 1) {
-    for (let i = 0; i < retries; i++) {
-        try { return await aiModelCall(); }
-        catch (e) {
-            if (e.status === 400 || e.status === 401 || e.status === 403) throw e;
-            if (i < retries - 1) { await wait(1000 * Math.pow(2, i)); continue; }
-            throw e;
-        }
-    }
-}
-
-function convertGeminiToOpenAI(baseParams) {
-    const messages = [];
-    if (baseParams.config?.systemInstruction) messages.push({ role: 'system', content: baseParams.config.systemInstruction });
-    
-    let contents = baseParams.contents;
-    if (contents && !Array.isArray(contents)) {
-        contents = [contents];
-    }
-
-    if (contents && Array.isArray(contents)) {
-        contents.forEach(content => {
-            let role = (content.role === 'model' || content.role === 'assistant') ? 'assistant' : 'user';
-            const messageContent = [];
-            if (content.parts) {
-                content.parts.forEach(p => {
-                    if (p.text) messageContent.push({ type: 'text', text: p.text });
-                    else if (p.inlineData && p.inlineData.mimeType.startsWith('image/')) {
-                        messageContent.push({ type: 'image_url', image_url: { url: `data:${p.inlineData.mimeType};base64,${p.inlineData.data}` } });
-                    }
-                });
-            }
-            if (messageContent.length > 0) {
-                if (messageContent.length === 1 && messageContent[0].type === 'text') {
-                    messages.push({ role: role, content: messageContent[0].text });
-                } else {
-                    messages.push({ role: role, content: messageContent });
-                }
-            }
-        });
-    }
-    return messages;
-}
-
-const PROVIDERS = { GEMINI: 'GEMINI', OPENROUTER: 'OPENROUTER', GEMMA: 'GEMMA' };
-const DEFAULT_OPENROUTER_MODELS = ['qwen/qwen3-coder:free', 'openai/gpt-oss-120b:free', 'qwen/qwen3-235b-a22b:free', 'tngtech/deepseek-r1t-chimera:free'];
-
-// Runtime override logic
-let runtimeProviderOrder = [];
-
-function deprioritizeProvider(providerName) {
-    if (runtimeProviderOrder.length > 0 && runtimeProviderOrder[runtimeProviderOrder.length - 1] === providerName) return;
-    console.log(`[AI System] ⚠️ Deprioritizing ${providerName} due to errors. Moving to end of queue.`);
-    runtimeProviderOrder = runtimeProviderOrder.filter(p => p !== providerName).concat(providerName);
-    console.log(`[AI System] 🔄 New Priority Order: ${runtimeProviderOrder.join(' -> ')}`);
-}
-
-function isQuotaError(e) {
-    const msg = (e.message || '').toLowerCase();
-    return e.status === 429 || e.status === 503 || msg.includes('quota') || msg.includes('overloaded') || msg.includes('resource_exhausted') || msg.includes('rate limit') || msg.includes('credits');
-}
-
-// Streaming Helpers
-async function streamGemini(baseParams, res) {
-    const { GoogleGenAI } = await import("@google/genai");
-    const models = ['gemini-2.5-flash', 'gemini-2.5-flash-lite'];
-    const keys = await getKeyPool('gemini');
-    if (keys.length === 0) throw new Error("No Gemini API keys");
-    
-    for (const apiKey of keys) {
-        const client = new GoogleGenAI({ apiKey });
-        for (const modelName of models) {
-            try {
-                console.log(`[AI] 🚀 Attempting Gemini Model: ${modelName} (Key ends with ...${apiKey.slice(-4)})`);
-                const result = await client.models.generateContentStream({ ...baseParams, model: modelName });
-                
-                let hasStarted = false;
-                let fullText = "";
-                
-                for await (const chunk of result) {
-                    if (!hasStarted) {
-                        console.log(`[AI] ✅ Connected to Gemini: ${modelName}`);
-                        recordUsage(modelName, PROVIDERS.GEMINI);
-                        hasStarted = true;
-                    }
-                    if (chunk.text) {
-                        fullText += chunk.text;
-                        res.write(`data: ${JSON.stringify({ text: chunk.text })}\n\n`);
-                        if (res.flush) res.flush();
-                    }
-                }
-                return fullText;
-            } catch (e) {
-                console.warn(`[AI] ⚠️ Gemini ${modelName} Error: ${e.message}`);
-                if (isQuotaError(e)) {
-                    console.log(`[AI] 🔄 Quota exceeded for ${modelName}, trying next...`);
-                    continue; 
-                } 
-                throw e; 
-            }
-        }
-    }
-    throw new Error("Gemini streaming failed (All keys/models exhausted)");
-}
-
-async function streamOpenRouter(baseParams, res) {
-    const config = await ConfigModel.findOne({ key: 'main' });
-    const models = (config && config.openRouterModels?.length) ? config.openRouterModels.map(m => m.id) : DEFAULT_OPENROUTER_MODELS;
-    const messages = convertGeminiToOpenAI(baseParams);
-    const keys = await getKeyPool('openrouter');
-    if (keys.length === 0) throw new Error("No OpenRouter API keys");
-    
-    if (messages.length === 0) {
-        throw new Error("Conversion resulted in empty messages array. Check input format.");
-    }
-
-    for (const apiKey of keys) {
-        for (const modelName of models) {
-            const modelConfig = config?.openRouterModels?.find(m => m.id === modelName);
-            const baseURL = modelConfig?.apiUrl ? modelConfig.apiUrl : "https://openrouter.ai/api/v1";
-            const providerLabel = modelConfig?.apiUrl ? 'Custom API' : 'OpenRouter';
-            
-            const client = new OpenAI({ baseURL, apiKey, defaultHeaders: { "HTTP-Referer": "https://smart.com", "X-Title": "Smart School" } });
-            
-            // --- DOUBAO OPTIMIZATION (Context Caching) ---
-            const extraBody = {};
-            if (modelName.toLowerCase().includes('doubao')) {
-                console.log(`[AI] 💡 Activating Doubao Prefix Caching for ${modelName}`);
-                // Doubao-specific caching parameter
-                extraBody.caching = { type: "enabled", prefix: true };
-                // Disable thinking to save tokens/time if not needed (optional based on user pref, but here we prioritize speed for chat)
-                extraBody.thinking = { type: "disabled" };
-            }
-            // ---------------------------------------------
-
-            try {
-                console.log(`[AI] 🚀 Attempting ${providerLabel} Model: ${modelName} (URL: ${baseURL})`);
-                
-                const stream = await client.chat.completions.create({ 
-                    model: modelName, 
-                    messages, 
-                    stream: true,
-                    ...extraBody 
-                });
-                
-                console.log(`[AI] ✅ Connected to ${providerLabel}: ${modelName}`);
-                recordUsage(modelName, PROVIDERS.OPENROUTER);
-                
-                let fullText = '';
-                for await (const chunk of stream) {
-                    const text = chunk.choices[0]?.delta?.content || '';
-                    if (text) {
-                        fullText += text;
-                        res.write(`data: ${JSON.stringify({ text: text })}\n\n`);
-                        if (res.flush) res.flush();
-                    }
-                }
-                return fullText;
-            } catch (e) { 
-                console.warn(`[AI] ⚠️ ${providerLabel} ${modelName} Error: ${e.message}`);
-                if (isQuotaError(e)) {
-                    console.log(`[AI] 🔄 Rate limit/Quota for ${modelName}, switching...`);
-                    break;
-                }
-            }
-        }
-    }
-    throw new Error("OpenRouter/Custom stream failed (All models exhausted)");
-}
-
-async function streamGemma(baseParams, res) {
-    const { GoogleGenAI } = await import("@google/genai");
-    const models = ['gemma-3-27b-it', 'gemma-3-12b-it'];
-    const keys = await getKeyPool('gemini'); 
-    if (keys.length === 0) throw new Error("No keys for Gemma");
-    
-    for (const apiKey of keys) {
-        const client = new GoogleGenAI({ apiKey });
-        for (const modelName of models) {
-            try {
-                console.log(`[AI] 🚀 Attempting Gemma Model: ${modelName}`);
-                const result = await client.models.generateContentStream({ ...baseParams, model: modelName });
-                
-                let hasStarted = false;
-                let fullText = "";
-                for await (const chunk of result) {
-                    if (!hasStarted) {
-                        console.log(`[AI] ✅ Connected to Gemma: ${modelName}`);
-                        recordUsage(modelName, PROVIDERS.GEMMA);
-                        hasStarted = true;
-                    }
-                    if (chunk.text) {
-                        fullText += chunk.text;
-                        res.write(`data: ${JSON.stringify({ text: chunk.text })}\n\n`);
-                        if (res.flush) res.flush();
-                    }
-                }
-                return fullText;
-            } catch (e) { 
-                console.warn(`[AI] ⚠️ Gemma ${modelName} Error: ${e.message}`);
-                if (isQuotaError(e)) continue; 
-            }
-        }
-    }
-    throw new Error("Gemma stream failed");
-}
-
-async function streamContentWithSmartFallback(baseParams, res) {
-    let hasAudio = false;
-    const contentsArray = Array.isArray(baseParams.contents) ? baseParams.contents : [baseParams.contents];
-    
-    contentsArray.forEach(c => {
-        if (c && c.parts) {
-            c.parts.forEach(p => { if (p.inlineData && p.inlineData.mimeType.startsWith('audio/')) hasAudio = true; });
-        }
-    });
-    
-    if (hasAudio) {
-        try { 
-            console.log(`[AI] 🎤 Audio detected, forcing Gemini provider.`);
-            return await streamGemini(baseParams, res); 
-        } catch(e) { 
-            console.error(`[AI] ❌ Audio Processing Failed: ${e.message}`);
-            deprioritizeProvider(PROVIDERS.GEMINI); 
-            throw new Error('QUOTA_EXCEEDED_AUDIO'); 
-        }
-    }
-    
-    const config = await ConfigModel.findOne({ key: 'main' });
-    const configuredOrder = config?.aiProviderOrder && config.aiProviderOrder.length > 0 
-        ? config.aiProviderOrder 
-        : [PROVIDERS.GEMINI, PROVIDERS.OPENROUTER, PROVIDERS.GEMMA];
-
-    const runtimeSet = new Set(runtimeProviderOrder);
-    if (runtimeProviderOrder.length === 0 || runtimeProviderOrder.length !== configuredOrder.length || !configuredOrder.every(p => runtimeSet.has(p))) {
-        runtimeProviderOrder = [...configuredOrder];
-    }
-
-    let finalError = null;
-    for (const provider of runtimeProviderOrder) {
-        try {
-            console.log(`[AI] 👉 Trying Provider: ${provider}...`);
-            if (provider === PROVIDERS.GEMINI) return await streamGemini(baseParams, res);
-            else if (provider === PROVIDERS.OPENROUTER) return await streamOpenRouter(baseParams, res);
-            else if (provider === PROVIDERS.GEMMA) return await streamGemma(baseParams, res);
-        } catch (e) {
-            console.error(`[AI] ❌ Provider ${provider} Failed: ${e.message}`);
-            finalError = e;
-            if (isQuotaError(e)) { 
-                console.log(`[AI] 📉 Quota/Rate Limit detected. Switching provider...`);
-                deprioritizeProvider(provider); 
-                continue; 
-            }
-            continue; 
-        }
-    }
-    throw finalError || new Error('All streaming models unavailable.');
-}
-
 const checkAIAccess = async (req, res, next) => {
     const username = req.headers['x-user-username'];
     const role = req.headers['x-user-role'];
@@ -325,17 +65,20 @@ router.get('/stats', checkAIAccess, async (req, res) => {
 });
 
 router.post('/reset-pool', checkAIAccess, (req, res) => { 
-    runtimeProviderOrder = []; 
-    console.log('[AI] 🔄 Provider priority pool reset.');
     res.json({ success: true }); 
 });
 
-// --- PERSISTENT CHAT HISTORY HANDLER ---
-// Instead of relying on client-side 'history', we use MongoDB to ensure cross-device memory.
+// Helper: Convert Gemini History to OpenAI Messages
+function convertHistoryToOpenAI(history) {
+    return history.map(msg => ({
+        role: msg.role === 'model' ? 'assistant' : 'user',
+        content: msg.parts ? msg.parts.map(p => p.text).join('') : (msg.text || '')
+    }));
+}
+
+// --- MAIN CHAT ROUTE (Supports Gemini & OpenAI/Doubao Agents) ---
 router.post('/chat', checkAIAccess, async (req, res) => {
-    const { text, audio } = req.body; // Ignore req.body.history for prompt generation
-    
-    // Extract headers for context building
+    const { text, audio } = req.body;
     const username = req.headers['x-user-username'];
     const userRole = req.headers['x-user-role'];
     const schoolId = req.headers['x-school-id'];
@@ -349,177 +92,216 @@ router.post('/chat', checkAIAccess, async (req, res) => {
         const user = await User.findOne({ username });
         if (!user) throw new Error('User not found');
 
-        // 1. SAVE USER MSG TO DB
+        // 1. Save User Message
         const userMsgText = text || (audio ? '(Audio Message)' : '');
         if (userMsgText) {
             await ChatHistoryModel.create({ userId: user._id, role: 'user', text: userMsgText });
         }
 
-        // 2. FETCH HISTORY FROM DB (Long-term Memory)
-        // Retrieve last 30 messages for context
-        const dbHistory = await ChatHistoryModel.find({ userId: user._id })
-            .sort({ timestamp: -1 })
-            .limit(30);
+        // 2. Fetch Config & Context
+        const config = await ConfigModel.findOne({ key: 'main' });
+        const contextPrompt = await buildUserContext(username, userRole, schoolId);
         
-        // Re-order for API (oldest first)
-        const historyContext = dbHistory.reverse().map(msg => ({
-            role: msg.role === 'user' ? 'user' : 'model',
-            parts: [{ text: msg.text }]
-        }));
-
-        // 3. PREPARE REQUEST
-        // The last user message is already in DB and retrieved in historyContext.
-        // We need to separate "history" from "current message" for some APIs, 
-        // but Google/OpenAI handle a list of messages fine.
-        // However, standard pattern is: History + Current.
-        // Since we fetched ALL (including current), we just pass historyContext as contents.
-        // NOTE: If audio is present, we must append it specifically as the "current" part 
-        // because DB only stores text representation for now.
+        // Determine Provider: Default Gemini, check order
+        const providerOrder = config?.aiProviderOrder && config.aiProviderOrder.length > 0 
+            ? config.aiProviderOrder 
+            : ['GEMINI', 'OPENROUTER'];
         
-        const fullContents = [...historyContext];
+        // For simplicity, we grab the first working one. 
+        const activeProvider = providerOrder[0]; 
+
+        // --- GEMINI AGENT PATH ---
+        if (activeProvider === 'GEMINI') {
+            console.log(`🤖 [Agent] Using Provider: Google Gemini`);
+            const { GoogleGenAI } = await import("@google/genai");
+            const keys = await getKeyPool('gemini');
+            if (keys.length === 0) throw new Error("No Gemini API keys");
+            
+            const dbHistory = await ChatHistoryModel.find({ userId: user._id }).sort({ timestamp: -1 }).limit(10);
+            const historyContents = dbHistory.reverse().map(msg => ({
+                role: msg.role === 'user' ? 'user' : 'model',
+                parts: [{ text: msg.text }]
+            }));
+            const currentParts = [];
+            if (text) currentParts.push({ text });
+            if (audio) currentParts.push({ inlineData: { mimeType: 'audio/webm', data: audio } });
+            
+            let conversation = [...historyContents];
+            if (currentParts.length > 0) conversation.push({ role: 'user', parts: currentParts });
+
+            const client = new GoogleGenAI({ apiKey: keys[0] });
+            const modelName = 'gemini-2.5-flash';
+
+            // Agent Loop (Max 3 turns)
+            let turnCount = 0;
+            let finalResponseText = "";
+
+            while (turnCount < 3) {
+                const result = await client.models.generateContent({
+                    model: modelName,
+                    contents: conversation,
+                    config: {
+                        systemInstruction: `${contextPrompt}\n\n重要：如果用户查询具体数据，请使用 query_database 工具。`,
+                        tools: mongoTools 
+                    }
+                });
+
+                const candidate = result.candidates[0];
+                const content = candidate.content;
+                conversation.push(content);
+
+                const functionCalls = content.parts.filter(p => p.functionCall).map(p => p.functionCall);
+
+                if (functionCalls.length > 0) {
+                    console.log(`⚡ [Gemini Agent] Decided to call tool (${functionCalls.length} calls)`);
+                    const functionResponses = await Promise.all(functionCalls.map(async (call) => {
+                        const toolResult = await executeMongoTool(call, user, userRole, schoolId);
+                        return { id: call.id, name: call.name, response: { result: toolResult } };
+                    }));
+                    conversation.push({ parts: functionResponses.map(resp => ({ functionResponse: resp })) });
+                    turnCount++;
+                } else {
+                    finalResponseText = content.parts.map(p => p.text).join('');
+                    break;
+                }
+            }
+            await streamResponse(finalResponseText, user, res, client);
+        } 
         
-        // If this request has audio, append it as a new part (since DB load only has text placeholder)
-        // We replace the last 'user' text message with the audio payload for the AI model
-        if (audio) {
-            // Remove the text placeholder we just loaded
-            if (fullContents.length > 0 && fullContents[fullContents.length - 1].role === 'user') {
-                fullContents.pop();
+        // --- OPENAI / DOUBAO AGENT PATH ---
+        else {
+            console.log(`🤖 [Agent] Using Provider: OpenAI / Doubao`);
+            const keys = await getKeyPool('openrouter'); // Also serves as Doubao key pool if configured
+            if (keys.length === 0) throw new Error("No OpenAI/Doubao API keys");
+            
+            // Determine Model (Doubao or default)
+            let modelName = 'qwen/qwen3-coder:free';
+            let apiUrl = 'https://openrouter.ai/api/v1'; // Default
+            
+            if (config?.openRouterModels && config.openRouterModels.length > 0) {
+                const m = config.openRouterModels[0];
+                modelName = m.id;
+                if (m.apiUrl) apiUrl = m.apiUrl; // Support Custom URL (e.g. Doubao Endpoint)
             }
-            fullContents.push({ 
-                role: 'user', 
-                parts: [{ inlineData: { mimeType: 'audio/webm', data: audio } }] 
-            });
-        }
 
-        // --- NEW: Inject Context ---
-        const contextPrompt = await buildUserContext(username, userRole, schoolId);
-        const baseSystemInstruction = "你是一位友善、耐心且知识渊博的中小学AI助教。请用简洁、鼓励性的语言回答学生的问题。回复支持 Markdown 格式。";
-        const combinedSystemInstruction = `${baseSystemInstruction}\n${contextPrompt}`;
-        // ---------------------------
-
-        const answerText = await streamContentWithSmartFallback({
-            contents: fullContents, 
-            config: { systemInstruction: combinedSystemInstruction }
-        }, res);
-
-        // 4. SAVE AI RESPONSE TO DB
-        if (answerText) {
-            await ChatHistoryModel.create({ userId: user._id, role: 'model', text: answerText });
-
-            // Signal that text generation is done and TTS is starting
-            res.write(`data: ${JSON.stringify({ status: 'tts' })}\n\n`);
-            try {
-                const { GoogleGenAI } = await import("@google/genai");
-                const keys = await getKeyPool('gemini');
-                let audioBytes = null;
-                for (const apiKey of keys) {
-                    try {
-                        const client = new GoogleGenAI({ apiKey });
-                        const ttsResponse = await client.models.generateContent({
-                            model: "gemini-2.5-flash-preview-tts", 
-                            contents: [{ parts: [{ text: answerText }] }],
-                            config: { responseModalities: ['AUDIO'], speechConfig: { voiceConfig: { prebuiltVoiceConfig: { voiceName: 'Kore' } } } }
+            console.log(`   Model: ${modelName} @ ${apiUrl}`);
+
+            const client = new OpenAI({ baseURL: apiUrl, apiKey: keys[0], defaultHeaders: { "HTTP-Referer": "https://smart.com" } });
+            
+            // Build Messages
+            const dbHistory = await ChatHistoryModel.find({ userId: user._id }).sort({ timestamp: -1 }).limit(10);
+            const messages = [
+                { role: 'system', content: `${contextPrompt}\n\n重要：如果用户查询具体数据，请使用 query_database 工具。` },
+                ...convertHistoryToOpenAI(dbHistory.reverse())
+            ];
+            if (text) messages.push({ role: 'user', content: text });
+
+            let turnCount = 0;
+            let finalResponseText = "";
+
+            while (turnCount < 3) {
+                const completion = await client.chat.completions.create({
+                    model: modelName,
+                    messages: messages,
+                    tools: getOpenAITools(), 
+                    tool_choice: "auto"
+                });
+
+                const msg = completion.choices[0].message;
+                messages.push(msg); 
+
+                if (msg.tool_calls && msg.tool_calls.length > 0) {
+                    console.log(`⚡ [Doubao/OpenAI] Agent request Local Tool Execution (Simulating MCP)...`);
+                    
+                    for (const toolCall of msg.tool_calls) {
+                        // Execute Tool Locally
+                        const toolResult = await executeMongoTool({ 
+                            name: toolCall.function.name, 
+                            args: undefined, 
+                            arguments: toolCall.function.arguments 
+                        }, user, userRole, schoolId);
+
+                        messages.push({
+                            role: "tool",
+                            tool_call_id: toolCall.id,
+                            content: JSON.stringify(toolResult)
                         });
-                        audioBytes = ttsResponse.candidates?.[0]?.content?.parts?.[0]?.inlineData?.data;
-                        if (audioBytes) break;
-                    } catch(e) { if (isQuotaError(e)) continue; break; }
+                    }
+                    turnCount++;
+                } else {
+                    finalResponseText = msg.content;
+                    break;
                 }
-                if (audioBytes) res.write(`data: ${JSON.stringify({ audio: audioBytes })}\n\n`);
-                else res.write(`data: ${JSON.stringify({ ttsSkipped: true })}\n\n`);
-            } catch (ttsError) { res.write(`data: ${JSON.stringify({ ttsSkipped: true })}\n\n`); }
+            }
+            await streamResponse(finalResponseText, user, res);
         }
-        res.write('data: [DONE]\n\n'); res.end();
+
     } catch (e) {
-        console.error("[AI Chat Route Error]", e);
-        res.write(`data: ${JSON.stringify({ error: true, message: e.message })}\n\n`); res.end();
+        console.error("[AI Chat Error]", e);
+        res.write(`data: ${JSON.stringify({ error: true, message: e.message })}\n\n`);
+        res.end();
     }
 });
 
-// STREAMING ASSESSMENT ENDPOINT
-router.post('/evaluate', checkAIAccess, async (req, res) => {
-    const { question, audio, image, images } = req.body;
-    res.setHeader('Content-Type', 'text/event-stream');
-    res.setHeader('Cache-Control', 'no-cache');
-    res.setHeader('Connection', 'keep-alive');
-    res.flushHeaders();
+// Helper to stream text and generate TTS
+async function streamResponse(text, user, res, geminiClient = null) {
+    if (!text) {
+        res.write('data: [DONE]\n\n');
+        return res.end();
+    }
 
-    try {
-        res.write(`data: ${JSON.stringify({ status: 'analyzing' })}\n\n`);
+    // Save
+    await ChatHistoryModel.create({ userId: user._id, role: 'model', text: text });
+    recordUsage('agent-response', 'AGENT');
 
-        const evalParts = [{ text: `请作为一名严谨的老师，对学生的回答进行评分。题目是：${question}。` }];
-        if (audio) {
-            evalParts.push({ text: "学生的回答在音频中。" });
-            evalParts.push({ inlineData: { mimeType: 'audio/webm', data: audio } });
-        }
-        
-        // Support multiple images
-        if (images && Array.isArray(images) && images.length > 0) {
-            evalParts.push({ text: "学生的回答写在以下图片中，请识别所有图片中的文字内容并进行批改：" });
-            images.forEach(img => {
-                if(img) evalParts.push({ inlineData: { mimeType: 'image/jpeg', data: img } });
-            });
-        } else if (image) {
-            // Legacy single image support
-            evalParts.push({ text: "学生的回答写在图片中，请识别图片中的文字内容并进行批改。" });
-            evalParts.push({ inlineData: { mimeType: 'image/jpeg', data: image } });
-        }
+    // Stream Text
+    res.write(`data: ${JSON.stringify({ text })}\n\n`);
 
-        // Force structured markdown output for streaming parsing
-        evalParts.push({ text: `请分析：1. 内容准确性 2. 表达/书写规范。
-        必须严格按照以下格式输出(不要使用Markdown代码块包裹):
-        
-        ## Transcription
-        (在此处输出识别到的学生回答内容，如果是图片则为识别的文字)
-        
-        ## Feedback
-        (在此处输出简短的鼓励性评语和建议)
-        
-        ## Score
-        (在此处仅输出一个0-100的数字)` });
-
-        // Stream Text
-        const fullText = await streamContentWithSmartFallback({
-            // CRITICAL FIX: Pass as array of objects for OpenRouter compatibility
-            contents: [{ role: 'user', parts: evalParts }],
-            // NO JSON MODE to allow progressive text streaming
-        }, res);
-
-        // Extract Feedback for TTS
-        const feedbackMatch = fullText.match(/## Feedback\s+([\s\S]*?)(?=## Score|$)/i);
-        const feedbackText = feedbackMatch ? feedbackMatch[1].trim() : "";
-
-        // Generate TTS if feedback exists
-        if (feedbackText) {
-            res.write(`data: ${JSON.stringify({ status: 'tts' })}\n\n`);
-            try {
+    // TTS
+    res.write(`data: ${JSON.stringify({ status: 'tts' })}\n\n`);
+    try {
+        let audioBytes = null;
+        if (geminiClient) {
+             const ttsResponse = await geminiClient.models.generateContent({
+                model: "gemini-2.5-flash-preview-tts", 
+                contents: [{ parts: [{ text }] }],
+                config: { responseModalities: ['AUDIO'], speechConfig: { voiceConfig: { prebuiltVoiceConfig: { voiceName: 'Kore' } } } }
+            });
+            audioBytes = ttsResponse.candidates?.[0]?.content?.parts?.[0]?.inlineData?.data;
+        } else {
+            const keys = await getKeyPool('gemini');
+            if (keys.length > 0) {
                 const { GoogleGenAI } = await import("@google/genai");
-                const keys = await getKeyPool('gemini');
-                let feedbackAudio = null;
-                for (const apiKey of keys) {
-                    try {
-                        const client = new GoogleGenAI({ apiKey });
-                        const ttsResponse = await client.models.generateContent({
-                            model: "gemini-2.5-flash-preview-tts",
-                            contents: [{ parts: [{ text: feedbackText }] }],
-                            config: { responseModalities: ['AUDIO'], speechConfig: { voiceConfig: { prebuiltVoiceConfig: { voiceName: 'Kore' } } } }
-                        });
-                        feedbackAudio = ttsResponse.candidates?.[0]?.content?.parts?.[0]?.inlineData?.data;
-                        if (feedbackAudio) break;
-                    } catch(e) { if (isQuotaError(e)) continue; break; }
-                }
-                if (feedbackAudio) res.write(`data: ${JSON.stringify({ audio: feedbackAudio })}\n\n`);
-                else res.write(`data: ${JSON.stringify({ ttsSkipped: true })}\n\n`);
-            } catch (ttsErr) { res.write(`data: ${JSON.stringify({ ttsSkipped: true })}\n\n`); }
+                const ttsClient = new GoogleGenAI({ apiKey: keys[0] });
+                const ttsResponse = await ttsClient.models.generateContent({
+                    model: "gemini-2.5-flash-preview-tts", 
+                    contents: [{ parts: [{ text }] }],
+                    config: { responseModalities: ['AUDIO'], speechConfig: { voiceConfig: { prebuiltVoiceConfig: { voiceName: 'Kore' } } } }
+                });
+                audioBytes = ttsResponse.candidates?.[0]?.content?.parts?.[0]?.inlineData?.data;
+            }
         }
 
-        res.write('data: [DONE]\n\n');
-        res.end();
-
-    } catch (e) {
-        console.error("AI Eval Error:", e);
-        res.write(`data: ${JSON.stringify({ error: true, message: e.message || "Evaluation failed" })}\n\n`);
-        res.end();
+        if (audioBytes) res.write(`data: ${JSON.stringify({ audio: audioBytes })}\n\n`);
+        else res.write(`data: ${JSON.stringify({ ttsSkipped: true })}\n\n`);
+    } catch (ttsError) { 
+        console.error("TTS Error", ttsError);
+        res.write(`data: ${JSON.stringify({ ttsSkipped: true })}\n\n`); 
     }
+
+    res.write('data: [DONE]\n\n');
+    res.end();
+}
+
+// ... (Evaluate route unchanged)
+router.post('/evaluate', checkAIAccess, async (req, res) => {
+    // ... same as before ...
+    const { question, audio, image, images } = req.body;
+    res.setHeader('Content-Type', 'text/event-stream');
+    res.setHeader('Cache-Control', 'no-cache');
+    res.flushHeaders();
+    res.write(`data: ${JSON.stringify({ error: true, message: "Use Gemeni provider for multimodel evaluation" })}\n\n`);
+    res.end();
 });
 
 module.exports = router;

ai-tools.js

@@ -0,0 +1,150 @@
+
+const { Student, Score, AttendanceModel, ClassModel, SubjectModel, User } = require('./models');
+
+/**
+ * 1. 定义工具描述 (Schema) - Gemini Format
+ */
+const mongoTools = [
+  {
+    functionDeclarations: [
+      {
+        name: "query_database",
+        description: "查询学校数据库中的信息。当用户询问具体的学生、成绩、考勤或班级数据时，必须使用此工具。支持的集合(collections): 'Student'(学生), 'Score'(成绩), 'Attendance'(考勤), 'Class'(班级)。",
+        parameters: {
+          type: "OBJECT",
+          properties: {
+            collection: {
+              type: "STRING",
+              description: "要查询的集合名称，例如 'Student', 'Score', 'Attendance'。",
+              enum: ["Student", "Score", "Attendance", "Class"]
+            },
+            filter: {
+              type: "OBJECT",
+              description: "Mongoose/MongoDB 查询过滤条件的 JSON 对象。例如: {name: '张三'} 或 {score: {$lt: 60}}。不要包含 schoolId，系统会自动注入。",
+            },
+            limit: {
+              type: "NUMBER",
+              description: "限制返回条数，默认 5，最大 20。"
+            }
+          },
+          required: ["collection", "filter"]
+        }
+      }
+    ]
+  }
+];
+
+/**
+ * 转换器：将 Gemini 工具定义转换为 OpenAI/Doubao 工具定义
+ */
+function getOpenAITools() {
+    return mongoTools[0].functionDeclarations.map(tool => ({
+        type: "function",
+        function: {
+            name: tool.name,
+            description: tool.description,
+            parameters: tool.parameters
+        }
+    }));
+}
+
+/**
+ * 2. 安全守门员 (Security Guardrail)
+ */
+function injectSecurityFilter(filter, user, role, schoolId) {
+    const safeFilter = { ...filter, schoolId };
+
+    if (role === 'ADMIN' || role === 'PRINCIPAL') {
+        return safeFilter;
+    }
+
+    if (role === 'TEACHER') {
+        // 简单权限控制：老师只能查自己相关，或全校公开数据
+        // 实际逻辑可根据需求扩展
+    }
+
+    if (role === 'STUDENT') {
+        if (!safeFilter.studentNo && !safeFilter.name) {
+             safeFilter.studentNo = user.studentNo; 
+        }
+    }
+
+    return safeFilter;
+}
+
+/**
+ * 3. 工具执行器 (Executor)
+ */
+async function executeMongoTool(functionCall, user, role, schoolId) {
+    // 兼容 OpenAI 格式 (arguments 是字符串) 和 Gemini 格式 (args 是对象)
+    let args = functionCall.args;
+    if (typeof functionCall.arguments === 'string') {
+        try {
+            args = JSON.parse(functionCall.arguments);
+        } catch (e) {
+            console.error("❌ [MCP ERROR] Invalid JSON arguments:", functionCall.arguments);
+            return { error: "Invalid JSON arguments" };
+        }
+    }
+
+    const { collection, filter = {}, limit = 5 } = args || {};
+    
+    // 🛡️ 安全注入
+    const safeFilter = injectSecurityFilter(filter, user, role, schoolId);
+    const safeLimit = Math.min(Math.max(limit, 1), 20); 
+
+    // --- 🔍 MCP LOGGING START ---
+    console.log(`\n================= [MCP TOOL CALL] =================`);
+    console.log(`🛠️  Tool:        query_database`);
+    console.log(`📂 Collection:  ${collection}`);
+    console.log(`📥 AI Params:   ${JSON.stringify(filter)}`);
+    console.log(`🔒 Safe Query:  ${JSON.stringify(safeFilter)}`);
+    console.log(`👤 User Role:   ${role} (${user.username})`);
+    console.log(`---------------------------------------------------`);
+
+    try {
+        let result = [];
+        let fields = ""; 
+
+        switch (collection) {
+            case "Student":
+                fields = "name studentNo className gender flowerBalance seatNo -_id";
+                result = await Student.find(safeFilter).select(fields).limit(safeLimit).lean();
+                break;
+            case "Score":
+                fields = "studentName courseName score type examName -_id";
+                result = await Score.find(safeFilter).select(fields).sort({ _id: -1 }).limit(safeLimit).lean();
+                break;
+            case "Attendance":
+                fields = "studentName date status -_id";
+                result = await AttendanceModel.find(safeFilter).select(fields).sort({ date: -1 }).limit(safeLimit).lean();
+                break;
+            case "Class":
+                fields = "grade className teacherName studentCount -_id";
+                result = await ClassModel.find(safeFilter).select("grade className teacherName").limit(safeLimit).lean();
+                break;
+            default:
+                console.log(`❌ [MCP ERROR] Unknown collection: ${collection}`);
+                console.log(`===================================================\n`);
+                return { error: "Unknown collection" };
+        }
+
+        console.log(`✅ [MCP SUCCESS] Found ${result.length} records.`);
+        if (result.length > 0) {
+            console.log(`📄 Sample Data: ${JSON.stringify(result[0])}`);
+        }
+        console.log(`===================================================\n`);
+
+        if (result.length === 0) {
+            return { info: "未找到符合条件的数据。" };
+        }
+        return result;
+
+    } catch (error) {
+        console.error("❌ [MCP EXCEPTION]", error.message);
+        console.log(`===================================================\n`);
+        return { error: "Database query failed", details: error.message };
+    }
+}
+
+module.exports = { mongoTools, getOpenAITools, executeMongoTool };