154 lines
3.9 KiB
JavaScript
154 lines
3.9 KiB
JavaScript
/**
|
||
* LLM Router · 模块入口（完整实现）
|
||
* GH-GMP-005 · M2 · Agent搬迁工程
|
||
*
|
||
* 大模型路由调度 · 通义千问API · 模型降级
|
||
*/
|
||
|
||
'use strict';
|
||
|
||
const QwenClient = require('./qwen-client');
|
||
|
||
const MODULE_NAME = 'llm-router';
|
||
const MODULE_VERSION = '1.0.0';
|
||
|
||
/**
|
||
* 模型路由配置
|
||
*/
|
||
const MODEL_ROUTES = {
|
||
reasoning: {
|
||
model: 'qwen-plus',
|
||
maxTokens: 4000,
|
||
temperature: 0.7,
|
||
description: '通用推理：工单分析、回执生成、讨论回复',
|
||
},
|
||
coding: {
|
||
model: 'qwen-coder-turbo',
|
||
maxTokens: 8000,
|
||
temperature: 0.3,
|
||
description: '代码生成：生成代码文件、代码审查',
|
||
},
|
||
simple: {
|
||
model: 'qwen-turbo',
|
||
maxTokens: 1000,
|
||
temperature: 0.5,
|
||
description: '简单任务：状态更新、格式化、短回复',
|
||
},
|
||
thinking: {
|
||
model: 'qwen-plus',
|
||
maxTokens: 8000,
|
||
temperature: 0.8,
|
||
description: '深度思考：架构设计、复杂决策',
|
||
},
|
||
};
|
||
|
||
/**
|
||
* 模型降级顺序
|
||
*/
|
||
const FALLBACK_ORDER = ['qwen-plus', 'qwen-turbo'];
|
||
|
||
let qwenClient = null;
|
||
let logger = console;
|
||
|
||
module.exports = {
|
||
name: MODULE_NAME,
|
||
version: MODULE_VERSION,
|
||
type: 'service',
|
||
MODEL_ROUTES,
|
||
|
||
async init(context) {
|
||
logger = (context && context.logger) || console;
|
||
logger.info(`[${MODULE_NAME}] 初始化中 v${MODULE_VERSION}...`);
|
||
|
||
qwenClient = new QwenClient({ logger });
|
||
|
||
const hasKey = !!process.env.GH_LLM_API_KEY;
|
||
logger.info(`[${MODULE_NAME}] API Key: ${hasKey ? '已配置 ✅' : '❌ 未配置'}`);
|
||
|
||
if (!hasKey) {
|
||
logger.warn(`[${MODULE_NAME}] 无API Key，LLM功能将不可用`);
|
||
}
|
||
|
||
return { status: hasKey ? 'ok' : 'warning', hasApiKey: hasKey };
|
||
},
|
||
|
||
async start(context) {
|
||
logger.info(`[${MODULE_NAME}] LLM路由就绪 · ${Object.keys(MODEL_ROUTES).length} 条路由`);
|
||
return { status: 'running' };
|
||
},
|
||
|
||
async stop() {
|
||
logger.info(`[${MODULE_NAME}] LLM路由停止`);
|
||
return { status: 'stopped' };
|
||
},
|
||
|
||
async healthCheck() {
|
||
return {
|
||
status: process.env.GH_LLM_API_KEY ? 'ok' : 'warning',
|
||
module: MODULE_NAME,
|
||
version: MODULE_VERSION,
|
||
routes: Object.keys(MODEL_ROUTES),
|
||
stats: qwenClient ? qwenClient.stats : null,
|
||
};
|
||
},
|
||
|
||
// ─── 对外接口 ───
|
||
|
||
/**
|
||
* 按路由类型调用LLM
|
||
* @param {string} routeType - 'reasoning'|'coding'|'simple'|'thinking'
|
||
* @param {Array} messages - [{role, content}]
|
||
* @returns {Promise<{content: string, usage: object, model: string}>}
|
||
*/
|
||
async chat(routeType, messages) {
|
||
const route = MODEL_ROUTES[routeType];
|
||
if (!route) {
|
||
throw new Error(`[${MODULE_NAME}] 未知路由类型: ${routeType}`);
|
||
}
|
||
|
||
// 尝试首选模型
|
||
try {
|
||
return await qwenClient.chat({
|
||
model: route.model,
|
||
messages,
|
||
maxTokens: route.maxTokens,
|
||
temperature: route.temperature,
|
||
});
|
||
} catch (err) {
|
||
logger.warn(
|
||
`[${MODULE_NAME}] ${route.model} 调用失败 · 尝试降级 · ${err.message}`
|
||
);
|
||
|
||
// 降级到备选模型
|
||
for (const fallback of FALLBACK_ORDER) {
|
||
if (fallback === route.model) continue;
|
||
try {
|
||
logger.info(`[${MODULE_NAME}] 降级到 ${fallback}`);
|
||
return await qwenClient.chat({
|
||
model: fallback,
|
||
messages,
|
||
maxTokens: route.maxTokens,
|
||
temperature: route.temperature,
|
||
});
|
||
} catch (fallbackErr) {
|
||
logger.warn(`[${MODULE_NAME}] ${fallback} 也失败 · ${fallbackErr.message}`);
|
||
}
|
||
}
|
||
|
||
throw new Error(`[${MODULE_NAME}] 所有模型均不可用`);
|
||
}
|
||
},
|
||
|
||
/**
|
||
* 直接指定模型调用（不走路由）
|
||
*/
|
||
async chatDirect({ model, messages, maxTokens, temperature }) {
|
||
return qwenClient.chat({ model, messages, maxTokens, temperature });
|
||
},
|
||
|
||
/**
|
||
* 获取QwenClient实例
|
||
*/
|
||
getClient: () => qwenClient,
|
||
};
|