257 lines
7.9 KiB
JavaScript
257 lines
7.9 KiB
JavaScript
/**
|
||
* ═══════════════════════════════════════════════════════════
|
||
* ZY-MIRROR-AGENT · Step 3 · 自主评估 — LLM 推理引擎
|
||
* ═══════════════════════════════════════════════════════════
|
||
*
|
||
* 将 diff 结果送入 LLM 模型进行评估：
|
||
* - 是否影响我们的搜索索引？
|
||
* - 是否需要更新本地数据库 schema？
|
||
* - 是否有新书源我们应该纳入？
|
||
* - 输出：升级方案（含具体变更步骤）
|
||
*
|
||
* 守护: 铸渊 · ICE-GL-ZY001
|
||
* 版权: 国作登字-2026-A-00037559
|
||
* ═══════════════════════════════════════════════════════════
|
||
*/
|
||
|
||
'use strict';
|
||
|
||
const https = require('https');
|
||
const http = require('http');
|
||
const { MIRROR_CONFIG } = require('./config');
|
||
|
||
/**
|
||
* 调用 LLM API
|
||
*/
|
||
function callLLM(provider, messages, maxTokens = 2000) {
|
||
return new Promise((resolve, reject) => {
|
||
const config = provider === 'deepseek' ? MIRROR_CONFIG.llm.primary : MIRROR_CONFIG.llm.fallback;
|
||
|
||
if (!config.api_key) {
|
||
return reject(new Error(`${provider} API key not configured`));
|
||
}
|
||
|
||
const url = new URL(config.api_url);
|
||
const isHttps = url.protocol === 'https:';
|
||
const client = isHttps ? https : http;
|
||
|
||
let body;
|
||
const headers = { 'Content-Type': 'application/json' };
|
||
|
||
if (provider === 'claude') {
|
||
headers['x-api-key'] = config.api_key;
|
||
headers['anthropic-version'] = '2023-06-01';
|
||
body = JSON.stringify({
|
||
model: config.model,
|
||
max_tokens: maxTokens,
|
||
messages
|
||
});
|
||
} else {
|
||
// DeepSeek / OpenAI compatible
|
||
headers['Authorization'] = `Bearer ${config.api_key}`;
|
||
body = JSON.stringify({
|
||
model: config.model,
|
||
messages,
|
||
max_tokens: maxTokens,
|
||
temperature: 0.3
|
||
});
|
||
}
|
||
|
||
const options = {
|
||
hostname: url.hostname,
|
||
port: url.port || (isHttps ? 443 : 80),
|
||
path: url.pathname,
|
||
method: 'POST',
|
||
headers: {
|
||
...headers,
|
||
'Content-Length': Buffer.byteLength(body)
|
||
},
|
||
timeout: 60000
|
||
};
|
||
|
||
const req = client.request(options, (res) => {
|
||
let data = '';
|
||
res.setEncoding('utf8');
|
||
res.on('data', chunk => { data += chunk; });
|
||
res.on('end', () => {
|
||
if (res.statusCode >= 200 && res.statusCode < 300) {
|
||
try {
|
||
const parsed = JSON.parse(data);
|
||
// Extract response text based on provider
|
||
let text;
|
||
if (provider === 'claude') {
|
||
text = parsed.content?.[0]?.text || '';
|
||
} else {
|
||
text = parsed.choices?.[0]?.message?.content || '';
|
||
}
|
||
resolve({ text, raw: parsed });
|
||
} catch (err) {
|
||
reject(new Error(`Failed to parse LLM response: ${err.message}`));
|
||
}
|
||
} else {
|
||
reject(new Error(`LLM API ${res.statusCode}: ${data.slice(0, 300)}`));
|
||
}
|
||
});
|
||
});
|
||
|
||
req.on('error', reject);
|
||
req.on('timeout', () => { req.destroy(); reject(new Error('LLM request timeout')); });
|
||
req.write(body);
|
||
req.end();
|
||
});
|
||
}
|
||
|
||
/**
|
||
* 构建评估 prompt
|
||
*/
|
||
function buildEvaluationPrompt(diffs) {
|
||
const systemPrompt = `你是「镜鉴」——光湖智库系统的镜面监控 Agent。
|
||
你的职责是分析第三方书库数据源的变化，判断我方系统是否需要跟进升级。
|
||
|
||
你必须从以下维度评估：
|
||
1. 搜索索引影响：变化是否影响我方搜索结果的准确性？
|
||
2. 数据库结构：是否需要更新我方的书目索引格式？
|
||
3. 新书源纳入：是否有值得我方纳入的新书目？
|
||
4. 可用性风险：数据源离线是否影响我方服务？
|
||
5. 上游兼容性：上游版本更新是否导致 API 接口变化？
|
||
|
||
请输出 JSON 格式的评估结果：
|
||
{
|
||
"needs_upgrade": true/false,
|
||
"urgency": "none" | "low" | "medium" | "high" | "critical",
|
||
"summary": "一句话摘要",
|
||
"recommendations": [
|
||
{
|
||
"action": "动作描述",
|
||
"reason": "原因",
|
||
"scope": "影响范围",
|
||
"steps": ["具体步骤1", "具体步骤2"]
|
||
}
|
||
],
|
||
"rollback_plan": "回滚方案（如果升级失败）",
|
||
"risk_assessment": "风险评估"
|
||
}
|
||
|
||
只输出 JSON，不要其他文字。`;
|
||
|
||
const userMessage = `以下是本次镜像扫描的差异报告：
|
||
|
||
${JSON.stringify(diffs, null, 2)}
|
||
|
||
请评估这些变化并给出升级建议。`;
|
||
|
||
return [
|
||
{ role: 'system', content: systemPrompt },
|
||
{ role: 'user', content: userMessage }
|
||
];
|
||
}
|
||
|
||
/**
|
||
* 执行 LLM 评估
|
||
* 先尝试 primary (DeepSeek)，失败回退到 fallback (Claude)
|
||
*/
|
||
async function evaluate(diffs) {
|
||
if (!diffs || diffs.length === 0) {
|
||
return {
|
||
needs_upgrade: false,
|
||
urgency: 'none',
|
||
summary: '无差异，无需评估',
|
||
recommendations: [],
|
||
evaluated_at: new Date().toISOString(),
|
||
model_used: 'none'
|
||
};
|
||
}
|
||
|
||
// 过滤掉无变化的 diff
|
||
const significantDiffs = diffs.filter(d =>
|
||
d.severity !== 'none' && !d.changes.every(c => c.type === 'no_change')
|
||
);
|
||
|
||
if (significantDiffs.length === 0) {
|
||
return {
|
||
needs_upgrade: false,
|
||
urgency: 'none',
|
||
summary: '所有数据源无显著变化',
|
||
recommendations: [],
|
||
evaluated_at: new Date().toISOString(),
|
||
model_used: 'none'
|
||
};
|
||
}
|
||
|
||
const messages = buildEvaluationPrompt(significantDiffs);
|
||
|
||
// 尝试 primary
|
||
try {
|
||
const result = await callLLM('deepseek', messages);
|
||
const evaluation = parseEvaluation(result.text);
|
||
evaluation.evaluated_at = new Date().toISOString();
|
||
evaluation.model_used = MIRROR_CONFIG.llm.primary.model;
|
||
return evaluation;
|
||
} catch (primaryErr) {
|
||
// 回退到 fallback
|
||
try {
|
||
const result = await callLLM('claude', messages);
|
||
const evaluation = parseEvaluation(result.text);
|
||
evaluation.evaluated_at = new Date().toISOString();
|
||
evaluation.model_used = MIRROR_CONFIG.llm.fallback.model;
|
||
evaluation._primary_error = primaryErr.message;
|
||
return evaluation;
|
||
} catch (fallbackErr) {
|
||
// 两个模型都失败，返回保守评估
|
||
return {
|
||
needs_upgrade: significantDiffs.some(d => d.severity === 'critical' || d.severity === 'action_needed'),
|
||
urgency: significantDiffs.some(d => d.severity === 'critical') ? 'high' : 'low',
|
||
summary: `LLM 评估失败，基于规则判断: ${significantDiffs.length} 个数据源有变化`,
|
||
recommendations: significantDiffs.map(d => ({
|
||
action: `检查数据源 ${d.source_name}`,
|
||
reason: d.changes.map(c => c.description).join('; '),
|
||
scope: d.source_id,
|
||
steps: ['人工检查差异报告', '决定是否跟进']
|
||
})),
|
||
rollback_plan: '无需回滚（未执行自动升级）',
|
||
risk_assessment: '评估模型不可用，建议人工审核',
|
||
evaluated_at: new Date().toISOString(),
|
||
model_used: 'rule_based_fallback',
|
||
_primary_error: primaryErr.message,
|
||
_fallback_error: fallbackErr.message
|
||
};
|
||
}
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 解析 LLM 输出为结构化评估结果
|
||
*/
|
||
function parseEvaluation(text) {
|
||
// 尝试直接 JSON 解析
|
||
try {
|
||
return JSON.parse(text);
|
||
} catch {
|
||
// 尝试从 markdown code block 中提取
|
||
const match = text.match(/```(?:json)?\s*([\s\S]*?)```/);
|
||
if (match) {
|
||
try {
|
||
return JSON.parse(match[1]);
|
||
} catch {
|
||
// fall through
|
||
}
|
||
}
|
||
|
||
// 都失败了，返回原始文本
|
||
return {
|
||
needs_upgrade: false,
|
||
urgency: 'low',
|
||
summary: '评估结果解析失败，原始输出已保存',
|
||
recommendations: [],
|
||
raw_output: text.slice(0, 2000)
|
||
};
|
||
}
|
||
}
|
||
|
||
module.exports = {
|
||
evaluate,
|
||
callLLM,
|
||
buildEvaluationPrompt,
|
||
parseEvaluation
|
||
};
|