mirror of
https://github.com/Wx-2025/ST-Amily2-Chat-Optimisation.git
synced 2026-06-12 11:05:49 +00:00
1a4a10d42d
### 修复 - **翰林院（RAG）API Key 污染**： - 修复 `saveSettingsFromUI` 无差别遍历翰林院面板内全部 `[data-setting-key]` 输入（包含被 `profile-sync` 接管隐藏的字段），导致掩码占位符 `••••••••` 被当作真值写回 `settings.rerank.apiKey` / `settings.retrieval.apiKey`，URL / model 也被 Profile 值覆盖到 legacy 字段。修复后会跳过祖先带 `data-profile-hidden` 的输入 - `getRerankSettings` / `getEmbedRetrievalSettings` 同时加入防御性还原：识别历史污染留下的 `••••••••` 时归为空字符串，避免取消 Profile 分配后实际请求带占位符 token 被 401 ---
440 lines
16 KiB
JavaScript
440 lines
16 KiB
JavaScript
'use strict';
|
||
|
||
import {
|
||
extension_settings
|
||
} from '/scripts/extensions.js';
|
||
import {
|
||
buildGoogleEmbeddingRequest,
|
||
parseGoogleEmbeddingResponse,
|
||
buildGoogleEmbeddingApiUrl
|
||
} from './utils/googleAdapter.js';
|
||
import { getSlotProfile } from './api/api-resolver.js';
|
||
import { extensionName } from '../utils/settings.js';
|
||
|
||
const MODULE_NAME = 'hanlinyuan-rag-core';
|
||
const GOOGLE_API_BASE_URL = 'https://generativelanguage.googleapis.com';
|
||
|
||
// profile-sync 在 UI 隐藏字段时填入的掩码占位符（const MASKED_KEY = '••••••••'）。
|
||
// 历史上 saveSettingsFromUI 曾把这个占位符写回 settings.{rerank,retrieval}.apiKey，
|
||
// 导致取消 Profile 分配后实际请求带占位符 token 被 401。这里做防御性还原。
|
||
const PROFILE_MASKED_KEY = '••••••••';
|
||
function sanitizeMaskedKey(key) {
|
||
return key === PROFILE_MASKED_KEY ? '' : key;
|
||
}
|
||
|
||
function getSettings() {
|
||
const root = extension_settings[extensionName];
|
||
const nested = root && root[MODULE_NAME];
|
||
if (nested) return nested;
|
||
// 读侧兼容：若迁移尚未触发（极早期调用），回退至旧顶层位置，避免空配置。
|
||
const legacy = extension_settings[MODULE_NAME];
|
||
if (legacy) return legacy;
|
||
console.error('[翰林院-API] 无法获取设置，API调用可能失败。');
|
||
return { retrieval: {}, rerank: {} };
|
||
}
|
||
|
||
/**
|
||
* 获取 Embedding 配置，优先从 ragEmbed 槽位 Profile 读取。
|
||
* Profile 存在时映射为 custom endpoint，覆盖旧 settings。
|
||
*/
|
||
export async function getEmbedRetrievalSettings() {
|
||
const profile = await getSlotProfile('ragEmbed');
|
||
if (profile) {
|
||
return {
|
||
apiEndpoint: profile.provider === 'google' ? 'google_direct' : 'custom',
|
||
customApiUrl: profile.apiUrl,
|
||
apiKey: sanitizeMaskedKey(profile.apiKey ?? ''),
|
||
embeddingModel: profile.model,
|
||
batchSize: getSettings().retrieval?.batchSize ?? 5,
|
||
};
|
||
}
|
||
const fallback = getSettings().retrieval || {};
|
||
return { ...fallback, apiKey: sanitizeMaskedKey(fallback.apiKey ?? '') };
|
||
}
|
||
|
||
/**
|
||
* 获取 Rerank 配置，优先从 ragRerank 槽位 Profile 读取。
|
||
*/
|
||
export async function getRerankSettings() {
|
||
const profile = await getSlotProfile('ragRerank');
|
||
if (profile) {
|
||
const manualSettings = getSettings().rerank || {};
|
||
return {
|
||
url: profile.apiUrl,
|
||
apiKey: sanitizeMaskedKey(profile.apiKey ?? ''),
|
||
model: profile.model,
|
||
top_n: manualSettings.top_n ?? 10,
|
||
apiMode: manualSettings.apiMode ?? 'custom',
|
||
};
|
||
}
|
||
const fallback = getSettings().rerank || {};
|
||
return { ...fallback, apiKey: sanitizeMaskedKey(fallback.apiKey ?? '') };
|
||
}
|
||
|
||
function normalizeApiResponse(responseData) {
|
||
let data = responseData;
|
||
if (typeof data === 'string') {
|
||
try {
|
||
data = JSON.parse(data);
|
||
} catch (e) {
|
||
console.error(`[翰林院-API] API响应JSON解析失败:`, e);
|
||
return { error: { message: 'Invalid JSON response' } };
|
||
}
|
||
}
|
||
if (data && typeof data.data === 'object' && data.data !== null && !Array.isArray(data.data)) {
|
||
if (Object.hasOwn(data.data, 'data')) {
|
||
data = data.data;
|
||
}
|
||
}
|
||
if (data && data.data) { // for /v1/models
|
||
return { data: data.data };
|
||
}
|
||
if (data && data.error) {
|
||
return { error: data.error };
|
||
}
|
||
return data;
|
||
}
|
||
|
||
export function getSanitizedBaseUrl(rawApiUrl) {
|
||
let baseUrl = rawApiUrl.trim();
|
||
if (baseUrl.endsWith('/')) {
|
||
baseUrl = baseUrl.slice(0, -1);
|
||
}
|
||
if (baseUrl.endsWith('/v1')) {
|
||
baseUrl = baseUrl.slice(0, -3);
|
||
}
|
||
// 兼容处理 /embeddings
|
||
if (baseUrl.endsWith('/embeddings')) {
|
||
baseUrl = baseUrl.slice(0, -11);
|
||
}
|
||
return baseUrl;
|
||
}
|
||
|
||
export async function fetchEmbeddingModels(overrideSettings = null) {
|
||
const settings = overrideSettings || await getEmbedRetrievalSettings();
|
||
const { apiEndpoint, apiKey, customApiUrl } = settings;
|
||
|
||
let modelsUrl;
|
||
let headers = {};
|
||
let responseParser;
|
||
|
||
switch (apiEndpoint) {
|
||
case 'google_direct':
|
||
if (!apiKey) throw new Error("Google直连模式需要API Key。");
|
||
|
||
const fetchGoogleModels = async (version) => {
|
||
const url = `${GOOGLE_API_BASE_URL}/${version}/models`;
|
||
console.log(`[翰林院] 正在从 Google API (${version}) 获取模型列表: ${url}`);
|
||
const response = await fetch(url, {
|
||
headers: { 'x-goog-api-key': apiKey },
|
||
});
|
||
if (!response.ok) {
|
||
console.warn(`获取 Google API (${version}) 模型列表失败: ${response.status}`);
|
||
return [];
|
||
}
|
||
const json = await response.json();
|
||
if (!json.models || !Array.isArray(json.models)) {
|
||
return [];
|
||
}
|
||
return json.models
|
||
.filter(model => model.supportedGenerationMethods?.includes('embedContent') || model.supportedGenerationMethods?.includes('batchEmbedContents'))
|
||
.map(model => model.name.replace('models/', ''));
|
||
};
|
||
|
||
const [v1Models, v1betaModels] = await Promise.all([
|
||
fetchGoogleModels('v1'),
|
||
fetchGoogleModels('v1beta')
|
||
]);
|
||
|
||
const allModels = [...new Set([...v1Models, ...v1betaModels])].sort();
|
||
return allModels;
|
||
|
||
case 'custom':
|
||
if (!customApiUrl) throw new Error("自定义模式需要API URL。");
|
||
if (!apiKey) throw new Error("自定义模式需要API Key。");
|
||
const customBaseUrl = getSanitizedBaseUrl(customApiUrl);
|
||
modelsUrl = `${customBaseUrl}/v1/models`;
|
||
headers = getApiHeaders(settings); // 这些模式需要认证头
|
||
console.log(`[翰林院] 正在从 ${modelsUrl} 获取模型列表 (需要认证)...`);
|
||
responseParser = (json) => {
|
||
if (!json.data || !Array.isArray(json.data)) {
|
||
throw new Error("模型API的响应格式无效: 未找到 'data' 数组。");
|
||
}
|
||
return json.data.map(m => m.id).sort();
|
||
};
|
||
// 对于custom模式，需要继续执行下面的fetch
|
||
break;
|
||
|
||
case 'local_proxy':
|
||
default:
|
||
if (!customApiUrl) throw new Error("本地代理模式需要API URL。");
|
||
const proxyBaseUrl = getSanitizedBaseUrl(customApiUrl);
|
||
modelsUrl = `${proxyBaseUrl}/v1/models`;
|
||
// 本地代理通常不需要认证头
|
||
headers = { 'Content-Type': 'application/json' };
|
||
console.log(`[翰林院] 正在从 ${modelsUrl} 获取模型列表 (无需认证)...`);
|
||
responseParser = (json) => {
|
||
if (!json.data || !Array.isArray(json.data)) {
|
||
throw new Error("模型API的响应格式无效: 未找到 'data' 数组。");
|
||
}
|
||
return json.data.map(m => m.id).sort();
|
||
};
|
||
break;
|
||
}
|
||
|
||
// 注意：st_backend case 和 google_direct case 已经提前返回，不会执行到这里
|
||
if (!modelsUrl) {
|
||
// 这个分支理论上不应该被执行，因为所有case都处理了
|
||
throw new Error('无法确定获取模型的有效路径。');
|
||
}
|
||
|
||
const response = await fetch(modelsUrl, {
|
||
method: 'GET',
|
||
headers: headers,
|
||
});
|
||
|
||
if (!response.ok) {
|
||
const errorBody = await response.text();
|
||
throw new Error(`获取模型列表失败 (${response.status}): ${errorBody}`);
|
||
}
|
||
|
||
const data = await response.json();
|
||
return responseParser(data);
|
||
}
|
||
|
||
export function getRerankBaseUrl(rawApiUrl) {
|
||
let baseUrl = rawApiUrl.trim();
|
||
if (baseUrl.endsWith('/')) {
|
||
baseUrl = baseUrl.slice(0, -1);
|
||
}
|
||
if (baseUrl.endsWith('/v1')) {
|
||
baseUrl = baseUrl.slice(0, -3);
|
||
}
|
||
// 兼容处理 /rerank
|
||
if (baseUrl.endsWith('/rerank')) {
|
||
baseUrl = baseUrl.slice(0, -7);
|
||
}
|
||
return baseUrl;
|
||
}
|
||
|
||
export async function fetchRerankModels() {
|
||
const settings = await getRerankSettings();
|
||
const { url, apiKey, apiMode = 'custom' } = settings;
|
||
|
||
if (!url) {
|
||
throw new Error("Rerank API URL 未提供。");
|
||
}
|
||
if (apiMode === 'custom' && !apiKey) {
|
||
throw new Error("自定义模式下，Rerank API Key 未提供。");
|
||
}
|
||
|
||
const baseUrl = getRerankBaseUrl(url);
|
||
const modelsUrl = `${baseUrl}/v1/models`;
|
||
const headers = { 'Content-Type': 'application/json' };
|
||
|
||
if (apiMode === 'custom') {
|
||
headers['Authorization'] = `Bearer ${apiKey}`;
|
||
}
|
||
|
||
console.log(`[翰林院-Rerank] 正在从 ${modelsUrl} 获取模型列表 (模式: ${apiMode})...`);
|
||
|
||
const response = await fetch(modelsUrl, {
|
||
method: 'GET',
|
||
headers: headers,
|
||
});
|
||
|
||
if (!response.ok) {
|
||
const errorBody = await response.text();
|
||
throw new Error(`获取Rerank模型列表失败 (${response.status}): ${errorBody}`);
|
||
}
|
||
|
||
const data = await response.json();
|
||
|
||
if (!data.data || !Array.isArray(data.data)) {
|
||
throw new Error("Rerank模型API的响应格式无效: 未找到 'data' 数组。");
|
||
}
|
||
|
||
return data.data
|
||
.map(m => m.id)
|
||
.sort();
|
||
}
|
||
|
||
export async function executeRerank(query, documents, rerankSettings = null) {
|
||
const resolved = rerankSettings || await getRerankSettings();
|
||
const { url, apiKey, model, top_n, apiMode = 'custom' } = resolved;
|
||
|
||
if (!url) throw new Error("Rerank API URL 未提供。");
|
||
if (apiMode === 'custom' && !apiKey) throw new Error("自定义模式下，Rerank API Key 未提供。");
|
||
|
||
const baseUrl = getRerankBaseUrl(url);
|
||
const rerankUrl = `${baseUrl}/v1/rerank`;
|
||
const headers = { 'Content-Type': 'application/json' };
|
||
|
||
if (apiMode === 'custom') {
|
||
headers['Authorization'] = `Bearer ${apiKey}`;
|
||
}
|
||
|
||
const body = JSON.stringify({
|
||
query: query,
|
||
documents: documents,
|
||
model: model,
|
||
top_n: top_n,
|
||
});
|
||
|
||
console.log(`[翰林院-Rerank] 正在向 ${rerankUrl} 发送请求 (模式: ${apiMode})...`);
|
||
|
||
const response = await fetch(rerankUrl, {
|
||
method: 'POST',
|
||
headers: headers,
|
||
body: body,
|
||
});
|
||
|
||
if (!response.ok) {
|
||
throw new Error(`Rerank API 请求失败 (${response.status}): ${await response.text()}`);
|
||
}
|
||
|
||
return await response.json();
|
||
}
|
||
|
||
|
||
export function getApiEndpointUrl(raw = false, overrideRetrieval = null) {
|
||
const {
|
||
apiEndpoint,
|
||
customApiUrl
|
||
} = overrideRetrieval || getSettings().retrieval;
|
||
let url;
|
||
switch (apiEndpoint) {
|
||
case 'openai':
|
||
url = 'https://api.openai.com';
|
||
break;
|
||
case 'azure':
|
||
case 'custom':
|
||
url = customApiUrl;
|
||
break;
|
||
default:
|
||
url = 'https://api.openai.com';
|
||
break;
|
||
}
|
||
if (raw) {
|
||
return url;
|
||
}
|
||
// 默认情况下，返回拼接好 /v1/embeddings 的完整URL
|
||
return getSanitizedBaseUrl(url) + '/v1/embeddings';
|
||
}
|
||
|
||
export function getApiHeaders(overrideRetrieval = null) {
|
||
const headers = {
|
||
'Content-Type': 'application/json'
|
||
};
|
||
const {
|
||
apiKey,
|
||
apiEndpoint
|
||
} = overrideRetrieval || getSettings().retrieval;
|
||
switch (apiEndpoint) {
|
||
case 'openai':
|
||
case 'custom':
|
||
headers['Authorization'] = `Bearer ${apiKey}`;
|
||
break;
|
||
case 'azure':
|
||
headers['api-key'] = apiKey;
|
||
break;
|
||
}
|
||
return headers;
|
||
}
|
||
|
||
export async function getEmbeddings(texts, signal = null) {
|
||
const settings = await getEmbedRetrievalSettings();
|
||
const { apiEndpoint, customApiUrl, apiKey, embeddingModel, batchSize = 5 } = settings;
|
||
|
||
const allEmbeddings = [];
|
||
|
||
for (let i = 0; i < texts.length; i += batchSize) {
|
||
if (signal?.aborted) throw new Error('AbortError');
|
||
const batch = texts.slice(i, i + batchSize);
|
||
let batchEmbeddings = [];
|
||
|
||
switch (apiEndpoint) {
|
||
case 'google_direct':
|
||
console.log('[翰林院-API] 使用Google直连模式获取向量。');
|
||
if (!apiKey) throw new Error('Google直连模式需要API Key。');
|
||
|
||
// 使用适配器构建URL和请求体；Key 通过 x-goog-api-key 头传递避免 URL 泄露
|
||
const googleUrl = buildGoogleEmbeddingApiUrl(GOOGLE_API_BASE_URL, embeddingModel);
|
||
const googleBody = buildGoogleEmbeddingRequest(batch, embeddingModel);
|
||
|
||
console.log(`[翰林院-API] 发送到 Google API 的请求 URL: ${googleUrl}`);
|
||
console.log(`[翰林院-API] 发送到 Google API 的请求体:`, JSON.stringify(googleBody, null, 2));
|
||
|
||
const googleResponse = await fetch(googleUrl, {
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
'x-goog-api-key': apiKey,
|
||
},
|
||
body: JSON.stringify(googleBody),
|
||
signal: signal,
|
||
});
|
||
|
||
if (!googleResponse.ok) {
|
||
const errorText = await googleResponse.text();
|
||
console.error(`[翰林院-API] Google API 错误响应: ${errorText}`);
|
||
throw new Error(`Google API Error: ${googleResponse.status} ${errorText}`);
|
||
}
|
||
const googleData = await googleResponse.json();
|
||
console.log(`[翰林院-API] 从 Google API 收到的响应:`, JSON.stringify(googleData, null, 2));
|
||
|
||
// 使用适配器解析响应
|
||
batchEmbeddings = parseGoogleEmbeddingResponse(googleData, batch);
|
||
break;
|
||
|
||
case 'custom':
|
||
case 'local_proxy':
|
||
default:
|
||
console.log(`[翰林院-API] 使用前端直连模式 (${apiEndpoint}) 获取向量。`);
|
||
if (!apiKey && apiEndpoint === 'custom') {
|
||
// 本地代理可以没有key，但自定义通常需要
|
||
// throw new Error('自定义模式需要API Key。');
|
||
}
|
||
const url = getApiEndpointUrl(false, settings); // 使用已解析的 settings
|
||
const headers = getApiHeaders(settings); // 使用已解析的 settings
|
||
|
||
const response = await fetch(url, {
|
||
method: 'POST',
|
||
headers,
|
||
body: JSON.stringify({
|
||
input: batch,
|
||
model: embeddingModel
|
||
}),
|
||
signal: signal,
|
||
});
|
||
|
||
if (!response.ok) {
|
||
const errorText = await response.text();
|
||
throw new Error(`神力获取失败 ${response.status}: ${errorText}`);
|
||
}
|
||
const result = await response.json();
|
||
if (result.data && Array.isArray(result.data)) {
|
||
batchEmbeddings = result.data.map(item => item.embedding);
|
||
} else {
|
||
throw new Error('API返回的向量数据格式不正确。');
|
||
}
|
||
break;
|
||
}
|
||
|
||
if (batchEmbeddings.length !== batch.length) {
|
||
throw new Error('获取到的向量数量与发送的文本数量不匹配。');
|
||
}
|
||
allEmbeddings.push(...batchEmbeddings);
|
||
|
||
// 避免速率限制
|
||
if (i + batchSize < texts.length) {
|
||
await new Promise(resolve => setTimeout(resolve, 200));
|
||
}
|
||
}
|
||
return allEmbeddings;
|
||
}
|
||
|
||
export async function testApiConnection() {
|
||
await getEmbeddings(['测试连接']);
|
||
}
|