feat: 初步支持文本模型

This commit is contained in:
foxhui
2025-12-16 17:50:29 +08:00
Unverified
parent 1c4fc1a314
commit 92c0e3c420
12 changed files with 1014 additions and 33 deletions
+90
View File
@@ -0,0 +1,90 @@
import { parseRequest } from '../src/server/parseChat.js';
// Mock options
const mockOptions = {
tempDir: './data/temp',
imageLimit: 5,
backendName: 'test-backend',
resolveModelId: (id) => id === 'gpt-4-text' ? 'resolved-gpt-4' : null,
getImagePolicy: () => 'optional',
getModelType: (id) => id === 'gpt-4-text' ? 'text' : 'image',
requestId: 'test-req-id',
logger: { info: console.log, error: console.error, warn: console.warn, debug: console.log }
};
async function testTextParsing() {
console.log('--- Testing Text Model Parsing ---');
const requestData = {
model: 'gpt-4-text',
stream: true,
messages: [
{ role: 'system', content: 'You are a cat.' },
{ role: 'user', content: 'Hello' },
{ role: 'assistant', content: 'Meow' },
{
role: 'user', content: [
{ type: 'text', text: 'Look at this ' },
{ type: 'image_url', image_url: { url: 'https://example.com/image.png' } }
]
}
]
};
const result = await parseRequest(requestData, mockOptions);
if (result.success) {
console.log('Success!');
console.log('Prompt:\n' + result.data.prompt);
console.log('Image Paths:', result.data.imagePaths);
if (result.data.prompt.includes('=== 系统指令') && result.data.prompt.includes('User: Look at this')) {
console.log('✅ Text Parsing Verification PASSED');
} else {
console.error('❌ Text Parsing Verification FAILED');
}
} else {
console.error('Failed:', result.error);
}
}
async function testImageParsing() {
console.log('\n--- Testing Image Model Parsing (Legacy) ---');
const requestData = {
model: 'some-image-model',
stream: false,
messages: [
{ role: 'user', content: 'Draw a cat' }
]
};
const options = {
...mockOptions,
resolveModelId: (id) => 'resolved-image-model',
getModelType: () => 'image'
};
const result = await parseRequest(requestData, options);
if (result.success) {
console.log('Success!');
console.log('Prompt:', result.data.prompt);
if (result.data.prompt === 'Draw a cat') {
console.log('✅ Image Parsing Verification PASSED');
} else {
console.error('❌ Image Parsing Verification FAILED');
}
} else {
console.error('Failed:', result.error);
}
}
(async () => {
try {
await testTextParsing();
await testImageParsing();
} catch (e) {
console.error(e);
}
})();
+3 -5
View File
@@ -58,7 +58,8 @@ const {
TEMP_DIR,
resolveModelId,
getModels,
getImagePolicy
getImagePolicy,
getModelType
} = backend;
/** @type {number} 服务器端口 */
@@ -102,16 +103,13 @@ const queueManager = createQueueManager(
: null
}
);
/**
* 路由处理器:负责 API 路由分发和鉴权
*/
const handleRequest = createRouter({
authToken: AUTH_TOKEN,
backendName,
getModels,
resolveModelId,
getImagePolicy,
getModelType,
tempDir: TEMP_DIR,
imageLimit: IMAGE_LIMIT,
queueManager
+295
View File
@@ -0,0 +1,295 @@
/**
* @fileoverview LMArena 适配器
*/
import {
sleep,
safeClick,
pasteImages
} from '../../browser/utils.js';
import {
fillPrompt,
submit,
waitApiResponse,
normalizePageError,
normalizeHttpError,
moveMouseAway,
waitForInput,
gotoWithCheck
} from '../utils/index.js';
import { logger } from '../../utils/logger.js';
// --- 配置常量 ---
const TARGET_URL = 'https://lmarena.ai/c/new?mode=direct';
/**
* 执行生图任务
* @param {object} context - 浏览器上下文 { page, client }
* @param {string} prompt - 提示词
* @param {string[]} imgPaths - 图片路径数组
* @param {string} [modelId] - 指定的模型 ID (可选)
* @param {object} [meta={}] - 日志元数据
* @returns {Promise<{image?: string, text?: string, error?: string}>} 生成结果
*/
async function generateImage(context, prompt, imgPaths, modelId, meta = {}) {
const { page, config } = context;
const textareaSelector = 'textarea';
try {
logger.info('适配器', '开启新会话...', meta);
const gotoResult = await gotoWithCheck(page, TARGET_URL);
if (gotoResult.error) return gotoResult;
// 1. 等待输入框加载
await waitForInput(page, textareaSelector, { click: false });
await sleep(1500, 2500);
// 2. 上传图片 (uploadImages)
if (imgPaths && imgPaths.length > 0) {
await pasteImages(page, textareaSelector, imgPaths);
}
// 3. 填写提示词 (fillPrompt)
await safeClick(page, textareaSelector, { bias: 'input' });
await fillPrompt(page, textareaSelector, prompt, meta);
// 4. 配置请求拦截 (用于修改模型 ID)
await page.unroute('**/*').catch(() => { });
if (modelId) {
logger.debug('适配器', `准备拦截请求`, meta);
await page.route(url => url.href.includes('/nextjs-api/stream'), async (route) => {
const request = route.request();
if (request.method() !== 'POST') return route.continue();
try {
const postData = request.postDataJSON();
if (postData && postData.modelAId) {
logger.info('适配器', `已拦截请求并修改模型: ${postData.modelAId} -> ${modelId}`, meta);
postData.modelAId = modelId;
await route.continue({ postData: JSON.stringify(postData) });
return;
}
} catch (e) {
logger.error('适配器', '拦截处理异常', { ...meta, error: e.message });
}
await route.continue();
});
}
// 5. 提交表单 (submit)
logger.debug('适配器', '点击发送...', meta);
await submit(page, {
btnSelector: 'button[type="submit"]',
inputTarget: textareaSelector,
meta
});
logger.info('适配器', '等待生成结果...', meta);
// 6. 等待 API 响应 (waitApiResponse)
let response;
try {
response = await waitApiResponse(page, {
urlMatch: '/nextjs-api/stream',
method: 'POST',
timeout: 120000,
meta
});
} catch (e) {
// 使用公共错误处理
const pageError = normalizePageError(e, meta);
if (pageError) return pageError;
throw e;
}
// 7. 解析响应结果
const content = await response.text();
// 8. 检查 HTTP 错误
const httpError = normalizeHttpError(response, content);
if (httpError) {
logger.error('适配器', `请求生成时返回错误: ${httpError.error}`, meta);
return { error: `请求生成时返回错误: ${httpError.error}` };
}
// 9. 解析文本流
// 格式示例:
// a0:"Hello"
// a0:" World"
// d:{"finishReason":"stop"}
let fullText = '';
const lines = content.split('\n');
for (const line of lines) {
if (line.startsWith('a0:')) {
try {
// 尝试解析 JSON 字符串内容
// line.substring(3) 应该是 JSON 字符串,如 "Hello"
const textPart = JSON.parse(line.substring(3));
fullText += textPart;
} catch (e) {
// 如果解析失败,可能是原生文本或其他格式
logger.warn('适配器', `解析文本块失败: ${line}`, meta);
}
}
}
if (fullText) {
logger.info('适配器', `获取文本成功,长度: ${fullText.length}`, meta);
return { text: fullText };
} else {
logger.warn('适配器', '未解析到有效文本内容', { ...meta, preview: content.substring(0, 150) });
// 如果没解析到 a0,尝试直接返回原始内容防空
return { error: '未解析到有效文本内容' };
}
} catch (err) {
// 顶层错误处理
const pageError = normalizePageError(err, meta);
if (pageError) return pageError;
logger.error('适配器', '生成任务失败', { ...meta, error: err.message });
return { error: `生成任务失败: ${err.message}` };
} finally {
// 清理拦截器
if (modelId) await page.unroute('**/*').catch(() => { });
// 任务结束,将鼠标移至安全区域
await moveMouseAway(page);
}
}
/**
* 适配器 manifest
*/
export const manifest = {
id: 'lmarena_text',
displayName: 'LMArena Text',
// 入口 URL
getTargetUrl(config, workerConfig) {
return TARGET_URL;
},
// 模型列表(从 models.js 迁移)
models: [
{ id: 'claude-opus-4-5-20251101-thinking-32k', codeName: '019ab8b2-9bcf-79b5-9fb5-149a7c67b7c0', imagePolicy: 'forbidden', type: 'text' },
{ id: 'claude-opus-4-5-20251101', 'codeName': '019adbec-8396-71cc-87d5-b47f8431a6a6', 'imagePolicy': 'forbidden', "type": "text" },
{ id: 'gemini-3-pro', codeName: '019a98f7-afcd-779f-8dcb-856cc3b3f078', imagePolicy: 'optional', type: 'text' },
{ id: 'grok-4.1-thinking', codeName: '019a9389-a9d3-77a8-afbb-4fe4dd3d8630', imagePolicy: 'forbidden', type: 'text' },
{ id: 'grok-4.1', codeName: '019a9389-a4d8-748d-9939-b4640198302e', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gpt-5.1-high', codeName: '019a8548-a2b1-70ce-b1be-eba096d41f58', imagePolicy: 'optional', type: 'text' },
{ id: 'gemini-2.5-pro', codeName: '0199f060-b306-7e1f-aeae-0ebb4e3f1122', imagePolicy: 'optional', type: 'text' },
{ id: 'claude-sonnet-4-5-20250929-thinking-32k', codeName: 'b0ea1407-2f92-4515-b9cc-b22a6d6c14f2', imagePolicy: 'forbidden', type: 'text' },
{ id: 'claude-opus-4-1-20250805-thinking-16k', codeName: 'f1a2eb6f-fc30-4806-9e00-1efd0d73cbc4', imagePolicy: 'forbidden', type: 'text' },
{ id: 'claude-sonnet-4-5-20250929', codeName: '019a2d13-28a5-7205-908c-0a58de904617', imagePolicy: 'forbidden', type: 'text' },
{ id: 'claude-opus-4-1-20250805', codeName: '96ae95fd-b70d-49c3-91cc-b58c7da1090b', imagePolicy: 'forbidden', type: 'text' },
{ id: 'chatgpt-4o-latest-20250326', codeName: '0199c1e0-3720-742d-91c8-787788b0a19b', imagePolicy: 'optional', type: 'text' },
{ id: 'gpt-5.1', codeName: '019a7ebf-0f3f-7518-8899-fca13e32d9dc', imagePolicy: 'optional', type: 'text' },
{ id: 'gpt-5-high', codeName: '983bc566-b783-4d28-b24c-3c8b08eb1086', imagePolicy: 'optional', type: 'text' },
{ id: 'o3-2025-04-16', codeName: 'cb0f1e24-e8e9-4745-aabc-b926ffde7475', imagePolicy: 'optional', type: 'text' },
{ id: 'qwen3-max-preview', codeName: '812c93cc-5f88-4cff-b9ca-c11a26599b0e', imagePolicy: 'forbidden', type: 'text' },
{ id: 'grok-4-1-fast-reasoning', codeName: '019aa41a-0a13-714a-beb1-be4a918a4b56', imagePolicy: 'forbidden', type: 'text' },
{ id: 'ernie-5.0-preview-1103', codeName: '019a4ca9-720d-75f5-9012-883ce8ff61df', imagePolicy: 'forbidden', type: 'text' },
{ id: 'kimi-k2-thinking-turbo', codeName: '019a59bc-8bb8-7933-92eb-fe143770c211', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gpt-5-chat', codeName: '4b11c78c-08c8-461c-938e-5fc97d56a40d', imagePolicy: 'optional', type: 'text' },
{ id: 'glm-4.6', codeName: 'f595e6f1-6175-4880-a9eb-377e390819e4', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-max-2025-09-23', codeName: '98ad8b8b-12cd-46cd-98de-99edde7e03eb', imagePolicy: 'forbidden', type: 'text' },
{ id: 'claude-opus-4-20250514-thinking-16k', codeName: '3b5e9593-3dc0-4492-a3da-19784c4bde75', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-235b-a22b-instruct-2507', codeName: 'ee7cb86e-8601-4585-b1d0-7c7380f8f6f4', imagePolicy: 'forbidden', type: 'text' },
{ id: 'grok-4-fast-chat', codeName: 'grok-4-fast-chat', imagePolicy: 'forbidden', type: 'text' },
{ id: 'deepseek-v3.2-thinking', codeName: '019adb32-bb7a-77eb-882f-b8e3aaa2b2fd', imagePolicy: 'forbidden', type: 'text' },
{ id: 'kimi-k2-0905-preview', codeName: 'b88e983b-9459-473d-8bf1-753932f1679a', imagePolicy: 'forbidden', type: 'text' },
{ id: 'kimi-k2-0711-preview', codeName: '7a3626fc-4e64-4c9e-821f-b449a4b43b6a', imagePolicy: 'forbidden', type: 'text' },
{ id: 'deepseek-v3.2', codeName: '019adb32-b716-7591-9a2f-c6882973e340', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-vl-235b-a22b-instruct', codeName: '716aa8ca-d729-427f-93ab-9579e4a13e98', imagePolicy: 'optional', type: 'text' },
{ id: 'mistral-large-3', codeName: '019acbac-df7c-73dc-9716-ebe040daaa4e', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gpt-4.1-2025-04-14', codeName: '14e9311c-94d2-40c2-8c54-273947e208b0', imagePolicy: 'optional', type: 'text' },
{ id: 'claude-opus-4-20250514', codeName: 'ee116d12-64d6-48a8-88e5-b2d06325cdd2', imagePolicy: 'forbidden', type: 'text' },
{ id: 'mistral-medium-2508', codeName: '27035fb8-a25b-4ec9-8410-34be18328afd', imagePolicy: 'optional', type: 'text' },
{ id: 'grok-4-0709', codeName: 'b9edb8e9-4e98-49e7-8aaf-ae67e9797a11', imagePolicy: 'optional', type: 'text' },
{ id: 'glm-4.5', codeName: 'd079ef40-3b20-4c58-ab5e-243738dbada5', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gemini-2.5-flash', codeName: '0199f059-3877-7cfe-bc80-e01b1a4a83de', imagePolicy: 'optional', type: 'text' },
{ id: 'gemini-2.5-flash-preview-09-2025', codeName: 'fc700d46-c4c1-4fec-88b5-f086876ae0bb', imagePolicy: 'optional', type: 'text' },
{ id: 'claude-haiku-4-5-20251001', codeName: '0199e8e9-01ed-73e0-96ba-cf43b286bf10', imagePolicy: 'forbidden', type: 'text' },
{ id: 'grok-4-fast-reasoning', codeName: '19b3730a-0369-49ba-ad9c-09e7337937f0', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-next-80b-a3b-instruct', codeName: '351fe482-eb6c-4536-857b-909e16c0bf52', imagePolicy: 'forbidden', type: 'text' },
{ id: 'longcat-flash-chat', codeName: '6fcbe051-f521-4dc7-8986-c429eb6191bf', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-235b-a22b-no-thinking', codeName: '1a400d9a-f61c-4bc2-89b4-a9b7e77dff12', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-235b-a22b-thinking-2507', codeName: '16b8e53a-cc7b-4608-a29a-20d4dac77cf2', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-vl-235b-a22b-thinking', codeName: '03c511f5-0d35-4751-aae6-24f918b0d49e', imagePolicy: 'optional', type: 'text' },
{ id: 'gpt-5-mini-high', codeName: '5fd3caa8-fe4c-41a5-a22c-0025b58f4b42', imagePolicy: 'optional', type: 'text' },
{ id: 'deepseek-v3-0324', codeName: '2f5253e4-75be-473c-bcfc-baeb3df0f8ad', imagePolicy: 'forbidden', type: 'text' },
{ id: 'hunyuan-vision-1.5-thinking', codeName: '6a3a1e04-050e-4cb4-9052-b9ac4bec0c38', imagePolicy: 'optional', type: 'text' },
{ id: 'o4-mini-2025-04-16', codeName: 'f1102bbf-34ca-468f-a9fc-14bcf63f315b', imagePolicy: 'optional', type: 'text' },
{ id: 'claude-sonnet-4-20250514', codeName: 'ac44dd10-0666-451c-b824-386ccfea7bcc', imagePolicy: 'forbidden', type: 'text' },
{ id: 'claude-3-7-sonnet-20250219-thinking-32k', codeName: 'be98fcfd-345c-4ae1-9a82-a19123ebf1d2', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-coder-480b-a35b-instruct', codeName: 'af033cbd-ec6c-42cc-9afa-e227fc12efe8', imagePolicy: 'forbidden', type: 'text' },
{ id: 'hunyuan-t1-20250711', codeName: 'ba8c2392-4c47-42af-bfee-c6c057615a91', imagePolicy: 'forbidden', type: 'text' },
{ id: 'mistral-medium-2505', codeName: '27b9f8c6-3ee1-464a-9479-a8b3c2a48fd4', imagePolicy: 'optional', type: 'text' },
{ id: 'qwen3-30b-a3b-instruct-2507', codeName: 'a8d1d310-e485-4c50-8f27-4bff18292a99', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gpt-4.1-mini-2025-04-14', codeName: '6a5437a7-c786-467b-b701-17b0bc8c8231', imagePolicy: 'optional', type: 'text' },
{ id: 'gemini-2.5-flash-lite-preview-09-2025-no-thinking', codeName: '75555628-8c14-402a-8d6e-43c19cb40116', imagePolicy: 'optional', type: 'text' },
{ id: 'gemini-2.5-flash-lite-preview-06-17-thinking', codeName: '04ec9a17-c597-49df-acf0-963da275c246', imagePolicy: 'optional', type: 'text' },
{ id: 'qwen3-235b-a22b', codeName: '2595a594-fa54-4299-97cd-2d7380d21c80', imagePolicy: 'forbidden', type: 'text' },
{ id: 'claude-3-5-sonnet-20241022', codeName: 'f44e280a-7914-43ca-a25d-ecfcc5d48d09', imagePolicy: 'forbidden', type: 'text' },
{ id: 'claude-3-7-sonnet-20250219', codeName: 'c5a11495-081a-4dc6-8d9a-64a4fd6f7bbc', imagePolicy: 'forbidden', type: 'text' },
{ id: 'glm-4.5-air', codeName: '7bfb254a-5d32-4ce2-b6dc-2c7faf1d5fe8', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-next-80b-a3b-thinking', codeName: '73cf8705-98c8-4b75-8d04-e3746e1c1565', imagePolicy: 'forbidden', type: 'text' },
{ id: 'minimax-m1', codeName: '87e8d160-049e-4b4e-adc4-7f2511348539', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gemma-3-27b-it', codeName: '789e245f-eafe-4c72-b563-d135e93988fc', imagePolicy: 'optional', type: 'text' },
{ id: 'grok-3-mini-high', codeName: '149619f1-f1d5-45fd-a53e-7d790f156f20', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gemini-2.0-flash-001', codeName: '7a55108b-b997-4cff-a72f-5aa83beee918', imagePolicy: 'optional', type: 'text' },
{ id: 'grok-3-mini-beta', codeName: '7699c8d4-0742-42f9-a117-d10e84688dab', imagePolicy: 'forbidden', type: 'text' },
{ id: 'mistral-small-2506', codeName: 'bbad1d17-6aa5-4321-949c-d11fb6289241', imagePolicy: 'optional', type: 'text' },
{ id: 'gpt-oss-120b', codeName: '6ee9f901-17b5-4fbe-9cc2-13c16497c23b', imagePolicy: 'forbidden', type: 'text' },
{ id: 'glm-4.5v', codeName: '9dab0475-a0cc-4524-84a2-3fd25aa8c768', imagePolicy: 'optional', type: 'text' },
{ id: 'command-a-03-2025', codeName: '0f785ba1-efcb-472d-961e-69f7b251c7e3', imagePolicy: 'forbidden', type: 'text' },
{ id: 'amazon-nova-experimental-chat-10-20', codeName: '019a4c75-256c-790b-9088-4694cc63c507', imagePolicy: 'forbidden', type: 'text' },
{ id: 'intellect-3', codeName: '019aebfd-af0e-7f0c-8f0d-96c588e4cd3b', imagePolicy: 'forbidden', type: 'text' },
{ id: 'o3-mini', codeName: 'c680645e-efac-4a81-b0af-da16902b2541', imagePolicy: 'forbidden', type: 'text' },
{ id: 'ling-flash-2.0', codeName: '71f96ca9-4cf8-4be7-bac2-2231613930a6', imagePolicy: 'forbidden', type: 'text' },
{ id: 'minimax-m2', codeName: '019a27e0-e7d8-7b0b-877c-a2106c6eb87d', imagePolicy: 'forbidden', type: 'text' },
{ id: 'step-3', codeName: '1ea13a81-93a7-4804-bcdd-693cd72e302d', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gpt-5-nano-high', codeName: '2dc249b3-98da-44b4-8d1e-6666346a8012', imagePolicy: 'optional', type: 'text' },
{ id: 'nova-2-lite', codeName: '019ae300-83b7-7717-a1e0-31accd1ff6fa', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwq-32b', codeName: '885976d3-d178-48f5-a3f4-6e13e0718872', imagePolicy: 'forbidden', type: 'text' },
{ id: 'llama-4-maverick-17b-128e-instruct', codeName: 'b5ad3ab7-fc56-4ecd-8921-bd56b55c1159', imagePolicy: 'optional', type: 'text' },
{ id: 'qwen3-30b-a3b', codeName: '9a066f6a-7205-4325-8d0b-d81cc4b049c0', imagePolicy: 'forbidden', type: 'text' },
{ id: 'claude-3-5-haiku-20241022', codeName: 'claude-3-5-haiku-20241022', imagePolicy: 'forbidden', type: 'text' },
{ id: 'ring-flash-2.0', codeName: '11ad4114-c868-4fed-b6e7-d535dc9c62f8', imagePolicy: 'forbidden', type: 'text' },
{ id: 'llama-3.3-70b-instruct', codeName: 'dcbd7897-5a37-4a34-93f1-76a24c7bb028', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gemma-3n-e4b-it', codeName: '896a3848-ae03-4651-963b-7d8f54b61ae8', imagePolicy: 'forbidden', type: 'text' },
{ id: 'gpt-oss-20b', codeName: 'ec3beb4b-7229-4232-bab9-670ee52dd711', imagePolicy: 'forbidden', type: 'text' },
{ id: 'mercury', codeName: '019a6f77-e20d-7c1d-a7cd-8bd926e7395d', imagePolicy: 'forbidden', type: 'text' },
{ id: 'olmo-3-32b-think', codeName: '019ac2ef-27e1-769f-8258-d131f79e28ef', imagePolicy: 'forbidden', type: 'text' },
{ id: 'magistral-medium-2506', codeName: '6337f479-2fc8-4311-a76b-8c957765cd68', imagePolicy: 'forbidden', type: 'text' },
{ id: 'mistral-small-3.1-24b-instruct-2503', codeName: '69f5d38a-45f5-4d3a-9320-b866a4035ed9', imagePolicy: 'optional', type: 'text' },
{ id: 'ibm-granite-h-small', codeName: '4ddb69f5-391a-4f78-af92-7d7328c18ab1', imagePolicy: 'forbidden', type: 'text' },
{ id: 'qwen3-vl-8b-thinking', codeName: '0199e3d1-a308-77b9-a650-41453e8ef2fb', imagePolicy: 'optional', type: 'text' },
{ id: 'qwen3-vl-8b-instruct', codeName: '0199e3d1-a713-7de2-a5dd-a1583cad9532', imagePolicy: 'optional', type: 'text' },
{ id: 'amazon.nova-pro-v1:0', codeName: 'a14546b5-d78d-4cf6-bb61-ab5b8510a9d6', imagePolicy: 'optional', type: 'text' },
{ id: 'glm-4.6v', codeName: '019b151a-7c3b-72a2-8811-0bf9317c2ef5', imagePolicy: 'optional', type: 'text' },
{ id: 'gpt-5.2-high', codeName: '019b1448-dafa-7f92-90c3-50e159c2263c', imagePolicy: 'optional', type: 'text' },
{ id: 'gpt-5.2', codeName: '019b1448-d548-78f4-8b98-788d72cbd057', imagePolicy: 'foptional', type: 'text' },
{ id: 'glm-4.6v-flash', codeName: '019b1536-49c0-73b2-8d45-403b8571568d', imagePolicy: 'optional', type: 'text' },
{ id: 'qwen3-omni-flash', codeName: '0199c9dc-e157-7458-bd49-5942363be215', imagePolicy: 'optional', type: 'text' }
],
// 模型 ID 解析
resolveModelId(modelKey) {
const model = this.models.find(m => m.id === modelKey);
return model ? model.codeName : null;
},
// 无需导航处理器
navigationHandlers: [],
// 核心生图方法
generateImage
};
+12
View File
@@ -110,6 +110,18 @@ export function getBackend() {
return poolManager.getImagePolicy(modelKey);
},
/**
* 获取模型类型
* @param {string} modelKey - 模型 key
* @returns {string} 'text' | 'image'
*/
getModelType: (modelKey) => {
if (!poolManager) {
return 'image';
}
return poolManager.getModelType(modelKey);
},
/**
* 获取 Cookies
* @param {string} [workerName] - Worker 名称
+12
View File
@@ -249,6 +249,18 @@ export class PoolManager {
return 'optional';
}
/**
* 获取模型类型
*/
getModelType(modelKey) {
for (const worker of this.workers) {
if (worker.supports(modelKey)) {
return worker.getModelType(modelKey);
}
}
return 'image';
}
/**
* 获取指定实例的 Cookies
*/
+21
View File
@@ -421,6 +421,27 @@ export class Worker {
}
}
/**
* 获取模型类型
*/
getModelType(modelKey) {
if (this.type === 'merge') {
if (modelKey.includes('/')) {
const [specifiedType, actualModel] = modelKey.split('/', 2);
if (this.mergeTypes.includes(specifiedType)) {
return registry.getModelType(specifiedType, actualModel);
}
}
for (const type of this.mergeTypes) {
const realId = registry.resolveModelId(type, modelKey);
if (realId) return registry.getModelType(type, modelKey);
}
return 'image';
} else {
return registry.getModelType(this.type, modelKey);
}
}
/**
* 导航到监控页面(空闲时)
*/
+20 -2
View File
@@ -197,7 +197,8 @@ class AdapterRegistry {
object: 'model',
created: Math.floor(Date.now() / 1000),
owned_by: id,
image_policy: m.imagePolicy
image_policy: m.imagePolicy,
type: m.type || 'image' // Default to image if not specified
}));
return { object: 'list', data };
@@ -243,6 +244,22 @@ class AdapterRegistry {
return model?.imagePolicy || IMAGE_POLICY.OPTIONAL;
}
/**
* 获取模型的类型
* @param {string} adapterId - 适配器 ID
* @param {string} modelKey - 模型 key
* @returns {string} 'text' | 'image'
*/
getModelType(adapterId, modelKey) {
const adapter = this.getAdapter(adapterId);
if (!adapter || !adapter.models) {
return 'image';
}
const model = adapter.models.find(m => m.id === modelKey);
return model?.type || 'image';
}
/**
* 聚合所有适配器的模型列表
* @returns {object}
@@ -258,7 +275,8 @@ class AdapterRegistry {
object: 'model',
created: Math.floor(Date.now() / 1000),
owned_by: id,
image_policy: m.imagePolicy
image_policy: m.imagePolicy,
type: m.type || 'image'
});
}
}
+3
View File
@@ -28,6 +28,7 @@ function checkAuth(req, authToken) {
* @param {Function} context.getModels - 获取模型列表函数
* @param {Function} context.resolveModelId - 解析模型 ID 函数
* @param {Function} context.getImagePolicy - 获取图片策略函数
* @param {Function} context.getModelType - 获取模型类型函数
* @param {string} context.tempDir - 临时目录
* @param {number} context.imageLimit - 图片数量限制
* @param {object} context.queueManager - 队列管理器
@@ -40,6 +41,7 @@ export function createRouter(context) {
getModels,
resolveModelId,
getImagePolicy,
getModelType,
tempDir,
imageLimit,
queueManager
@@ -138,6 +140,7 @@ export function createRouter(context) {
backendName,
resolveModelId,
getImagePolicy,
getModelType,
requestId,
logger
});
+157 -20
View File
@@ -56,6 +56,7 @@ function parseError(code, customMessage) {
* @param {string} options.backendName - 后端名称
* @param {Function} options.resolveModelId - 模型 ID 解析函数
* @param {Function} options.getImagePolicy - 获取图片策略函数
* @param {Function} options.getModelType - 获取模型类型函数
* @param {string} options.requestId - 请求 ID
* @param {Function} options.logger - 日志函数
* @returns {Promise<ParseResult>} 解析结果
@@ -67,6 +68,7 @@ export async function parseRequest(data, options) {
backendName,
resolveModelId,
getImagePolicy,
getModelType,
requestId,
logger
} = options;
@@ -79,6 +81,156 @@ export async function parseRequest(data, options) {
return parseError(ERROR_CODES.NO_MESSAGES);
}
// 1. 解析模型参数与类型
let modelKey = null;
let isTextMode = false;
if (data.model) {
const resolved = resolveModelId(data.model);
if (resolved) {
modelKey = data.model;
logger.info('服务器', `触发模型: ${data.model}`, { id: requestId });
// 判定是否为文本模式
const type = getModelType ? getModelType(data.model) : 'image';
isTextMode = type === 'text';
if (isTextMode) {
logger.info('服务器', '解析模式: 文本对话 (虚拟上下文构建)', { id: requestId });
} else {
logger.info('服务器', '解析模式: 图像生成 (仅取最后一条)', { id: requestId });
}
} else {
return parseError(ERROR_CODES.INVALID_MODEL, `模型无效/后端 ${backendName} 不支持: ${data.model}`);
}
} else {
logger.info('服务器', '未指定模型,使用网页默认', { id: requestId });
}
// ============================================================
// 分支 A: 文本模型解析 (构建虚拟上下文)
// ============================================================
if (isTextMode) {
return await parseTextRequest(messages, tempDir, imageLimit, modelKey, isStreaming);
}
// ============================================================
// 分支 B: 生图模型解析 (原有逻辑)
// ============================================================
return await parseImageRequest(messages, tempDir, imageLimit, modelKey, isStreaming, getImagePolicy);
}
/**
* 解析文本请求 (构建虚拟上下文)
*/
async function parseTextRequest(messages, tempDir, imageLimit, modelId, isStreaming) {
let systemPrompt = '';
let historyPrompt = '';
let currentPrompt = '';
const imagePaths = [];
let globalImageCount = 0;
// 辅助函数:处理单条消息内容
async function processContent(content) {
let textBuffer = '';
if (typeof content === 'string') {
textBuffer += content;
} else if (Array.isArray(content)) {
for (const item of content) {
if (item.type === 'text') {
textBuffer += item.text;
} else if (item.type === 'image_url' && item.image_url?.url) {
globalImageCount++;
// 图片数量限制检查
if (imageLimit > 0 && globalImageCount > imageLimit) {
textBuffer += `[图片${globalImageCount} (已忽略:超过限制)]`;
continue;
}
const url = item.image_url.url;
if (url.startsWith('data:image')) {
const imagePath = await saveBase64Image(url, tempDir);
if (imagePath) {
imagePaths.push(imagePath);
// 插入占位符
textBuffer += `[图片${globalImageCount}]`;
} else {
textBuffer += `[图片${globalImageCount} (上传失败)]`;
}
} else {
textBuffer += `[图片${globalImageCount} (无效链接)]`;
}
}
}
}
return textBuffer;
}
// 1. 提取 System Prompt
const systemMsg = messages.find(m => m.role === 'system');
if (systemMsg) {
const content = await processContent(systemMsg.content);
if (content) {
systemPrompt = `=== 系统指令 (永远置顶) ===\n${content}\n\n`;
}
}
// 2. 区分历史和当前消息
// 找到最后一条 user 消息的索引
let lastUserIndex = -1;
for (let i = messages.length - 1; i >= 0; i--) {
if (messages[i].role === 'user') {
lastUserIndex = i;
break;
}
}
if (lastUserIndex === -1) {
return parseError(ERROR_CODES.NO_USER_MESSAGES);
}
// 3. 构建历史对话 (不包含 system 和 最后一条 user)
const historyMessages = messages.filter((m, index) => {
return m.role !== 'system' && index < lastUserIndex;
});
if (historyMessages.length > 0) {
historyPrompt += `=== 历史对话 (滑动窗口或摘要) ===\n`;
for (const msg of historyMessages) {
const roleName = msg.role === 'user' ? 'User' : 'AI';
const content = await processContent(msg.content);
historyPrompt += `${roleName}: ${content}\n`;
}
historyPrompt += `\n`;
}
// 4. 构建当前输入
const lastUserMsg = messages[lastUserIndex];
const currentContent = await processContent(lastUserMsg.content);
currentPrompt = `=== 当前输入 ===\nUser: ${currentContent}`;
// 5. 合并最终 Prompt
const finalPrompt = systemPrompt + historyPrompt + currentPrompt;
return {
success: true,
data: {
prompt: finalPrompt,
imagePaths,
modelId,
modelName: modelId,
isStreaming
}
};
}
/**
* 解析生图请求 (原有逻辑)
*/
async function parseImageRequest(messages, tempDir, imageLimit, modelId, isStreaming, getImagePolicy) {
// 筛选用户消息
const userMessages = messages.filter(m => m.role === 'user');
if (userMessages.length === 0) {
@@ -127,31 +279,16 @@ export async function parseRequest(data, options) {
prompt = prompt.trim();
// 解析模型参数
let modelKey = null;
if (data.model) {
// 只校验模型是否支持,不解析
const resolved = resolveModelId(data.model);
if (resolved) {
modelKey = data.model; // 保留原始 modelKey,由 PoolManager 自行解析
logger.info('服务器', `触发模型: ${data.model}`, { id: requestId });
} else {
return parseError(ERROR_CODES.INVALID_MODEL, `模型无效/后端 ${backendName} 不支持: ${data.model}`);
}
} else {
logger.info('服务器', '未指定模型,使用网页默认', { id: requestId });
}
// 图片策略校验
const hasImage = imagePaths.length > 0;
const policy = data.model ? getImagePolicy(data.model) : IMAGE_POLICY.OPTIONAL;
const policy = modelId ? getImagePolicy(modelId) : IMAGE_POLICY.OPTIONAL;
if (policy === IMAGE_POLICY.REQUIRED && !hasImage) {
return parseError(ERROR_CODES.IMAGE_REQUIRED, `模型 ${data.model} 需要参考图`);
return parseError(ERROR_CODES.IMAGE_REQUIRED, `模型 ${modelId} 需要参考图`);
}
if (policy === IMAGE_POLICY.FORBIDDEN && hasImage) {
return parseError(ERROR_CODES.IMAGE_FORBIDDEN, `模型 ${data.model} 不支持图片输入`);
return parseError(ERROR_CODES.IMAGE_FORBIDDEN, `模型 ${modelId} 不支持图片输入`);
}
return {
@@ -159,8 +296,8 @@ export async function parseRequest(data, options) {
data: {
prompt,
imagePaths,
modelId: modelKey, // 返回原始 modelKey
modelName: data.model || null,
modelId,
modelName: modelId,
isStreaming
}
};
+3
View File
@@ -135,13 +135,16 @@ export function createQueueManager(queueConfig, callbacks) {
}
// 发送成功响应
logger.info('服务器', '准备发送响应...', { id, isStreaming, contentLength: finalContent.length });
if (isStreaming) {
const chunk = buildChatCompletionChunk(finalContent, modelName);
sendSse(res, chunk);
sendSseDone(res);
logger.info('服务器', '流式响应已结束', { id });
} else {
const response = buildChatCompletion(finalContent, modelName);
sendJson(res, 200, response);
logger.info('服务器', 'JSON 响应已发送', { id });
}
} catch (err) {
+3 -6
View File
@@ -19,9 +19,6 @@ const EXAMPLE_CONFIG_PATH = path.join(process.cwd(), 'config.example.yaml');
// 模块级缓存:确保配置只从磁盘读取一次
let cachedConfig = null;
// 有效的适配器类型
const VALID_ADAPTER_TYPES = ['lmarena', 'gemini', 'gemini_biz', 'nanobananafree_ai', 'zai_is', 'merge'];
/**
* 解析用户数据目录路径
* @param {string|undefined} userDataMark - 用户数据标记
@@ -84,9 +81,9 @@ function validateWorker(worker, instanceName, index) {
if (!worker.type) {
throw new Error(`instances[${instanceName}].workers[${index}] (${worker.name}) 缺少必需字段: type`);
}
if (!VALID_ADAPTER_TYPES.includes(worker.type)) {
throw new Error(`Worker "${worker.name}" 的 type "${worker.type}" 无效。有效值: ${VALID_ADAPTER_TYPES.join(', ')}`);
}
// 移除对 type 的硬编码校验,允许动态加载新适配器
// if (!VALID_ADAPTER_TYPES.includes(worker.type)) { ... }
if (worker.type === 'merge') {
if (!worker.mergeTypes || !Array.isArray(worker.mergeTypes) || worker.mergeTypes.length === 0) {
throw new Error(`Worker "${worker.name}" 类型为 merge,但缺少有效的 mergeTypes 数组`);
+395
View File
@@ -0,0 +1,395 @@
这是一个非常好的架构升级思路。为了支持文本模型(Chat Model)并构建“虚拟上下文”,我们需要重构 `src/server/parseChat.js`,让它具备根据模型类型(图片 vs 文本)采用不同解析策略的能力。
以下是具体的修改方案:
### 修改思路
1. **区分模型类型**:我们需要知道当前请求的模型是“生图模型”还是“文本模型”。
* *方案*:在 `registry.js` 中为模型定义增加 `type` 字段(如 `type: 'text'`),或者在 `parseChat.js` 中简单地通过判断模型 ID 是否包含特定关键字(如 `gpt-4`, `claude`)来区分。为了演示方便,我将在代码中添加一个简单的判断逻辑。
2. **文本模型解析逻辑**
* 不再只取最后一条消息。
* 遍历 `messages` 数组。
* 提取 `system` 消息作为“系统指令”。
* 提取之前的 `user``assistant` 消息作为“历史对话”。
* 提取最后一条 `user` 消息作为“当前输入”。
* **图片处理**:在遍历过程中,遇到图片不仅要保存文件,还要在文本中插入 `[图片1]` 占位符,并将图片路径收集到 `imagePaths` 数组中。
### 修改代码
你需要修改 `src/server/parseChat.js`
#### 1\. 新增辅助函数 `buildTextContext`
这个函数负责将 OpenAI 格式的 `messages` 数组转换为你要求的“虚拟上下文”字符串格式。
#### 2\. 改造 `parseRequest` 主函数
增加分支逻辑:如果是文本模型走新逻辑,如果是生图模型走原有逻辑(只取最后一条)。
以下是完整的修改后的 `src/server/parseChat.js` 代码:
```javascript
/**
* @fileoverview 请求解析模块
* @description 负责解析聊天请求、提取提示词和处理图片
*/
import fs from 'fs';
import path from 'path';
import sharp from 'sharp';
import { IMAGE_POLICY } from '../backend/registry.js';
import { ERROR_CODES, getErrorMessage } from './errors.js';
/**
* 构造解析错误结果
*/
function parseError(code, customMessage) {
return {
success: false,
error: {
code,
error: customMessage || getErrorMessage(code)
}
};
}
/**
* 判断是否为文本模型 (根据模型ID特征简单判断,你可以根据 registry 里的配置来判断)
* @param {string} modelId
* @returns {boolean}
*/
function isTextModel(modelId) {
if (!modelId) return false;
// 这里可以根据实际情况修改,比如检查 registry.getModelType(modelId)
// 示例:如果 ID 包含 gpt, claude, llama, qwen 且不是 image 模型,则认为是文本模型
// 现有的图片模型通常有 image 关键字,或者在 registry 中明确定义。
// 为了简单起见,假设除了已知的生图模型外,其他都是文本模型,或者你可以维护一个列表。
const textKeywords = ['gpt-4', 'claude', 'llama', 'qwen', 'deepseek', 'gemini-pro'];
// 注意:gemini-3-pro-image-preview 是生图的,需排除
if (modelId.includes('image') || modelId.includes('flux') || modelId.includes('seedream')) return false;
return textKeywords.some(k => modelId.includes(k));
}
/**
* 解析聊天请求
*/
export async function parseRequest(data, options) {
const {
tempDir,
imageLimit,
backendName,
resolveModelId,
getImagePolicy,
requestId,
logger
} = options;
const messages = data.messages;
const isStreaming = data.stream === true;
// 验证 messages
if (!messages || messages.length === 0) {
return parseError(ERROR_CODES.NO_MESSAGES);
}
// 1. 解析模型参数
let modelKey = null;
let isTextMode = false;
if (data.model) {
const resolved = resolveModelId(data.model);
if (resolved) {
modelKey = data.model;
logger.info('服务器', `触发模型: ${data.model}`, { id: requestId });
// 判定是否为文本模式 (你需要根据实际情况实现 isTextModel 或传入 getModelType)
// 这里假设如果不包含特定生图关键词就是文本模型,或者你可以扩展 options 传入判断函数
isTextMode = isTextModel(data.model);
if (isTextMode) {
logger.info('服务器', '解析模式: 文本对话 (虚拟上下文构建)', { id: requestId });
} else {
logger.info('服务器', '解析模式: 图像生成 (仅取最后一条)', { id: requestId });
}
} else {
return parseError(ERROR_CODES.INVALID_MODEL, `模型无效/后端 ${backendName} 不支持: ${data.model}`);
}
} else {
logger.info('服务器', '未指定模型,使用网页默认', { id: requestId });
}
// ============================================================
// 分支 A: 文本模型解析 (构建虚拟上下文)
// ============================================================
if (isTextMode) {
return await parseTextRequest(messages, tempDir, imageLimit, modelKey, isStreaming);
}
// ============================================================
// 分支 B: 生图模型解析 (原有逻辑,只取最后一条 User 消息)
// ============================================================
return await parseImageRequest(messages, tempDir, imageLimit, modelKey, isStreaming, getImagePolicy);
}
/**
* 解析文本请求 (构建虚拟上下文)
*/
async function parseTextRequest(messages, tempDir, imageLimit, modelId, isStreaming) {
let systemPrompt = '';
let historyPrompt = '';
let currentPrompt = '';
const imagePaths = [];
let globalImageCount = 0;
// 辅助函数:处理单条消息内容
async function processContent(content) {
let textBuffer = '';
if (typeof content === 'string') {
textBuffer += content;
} else if (Array.isArray(content)) {
for (const item of content) {
if (item.type === 'text') {
textBuffer += item.text;
} else if (item.type === 'image_url' && item.image_url?.url) {
globalImageCount++;
// 图片数量限制检查
if (imageLimit > 0 && globalImageCount > imageLimit) {
textBuffer += `[图片${globalImageCount} (已忽略:超过限制)]`;
continue;
}
const url = item.image_url.url;
if (url.startsWith('data:image')) {
const imagePath = await saveBase64Image(url, tempDir);
if (imagePath) {
imagePaths.push(imagePath);
// 插入占位符
textBuffer += `[图片${globalImageCount}]`;
} else {
textBuffer += `[图片${globalImageCount} (上传失败)]`;
}
} else {
textBuffer += `[图片${globalImageCount} (无效链接)]`;
}
}
}
}
return textBuffer;
}
// 1. 提取 System Prompt
const systemMsg = messages.find(m => m.role === 'system');
if (systemMsg) {
const content = await processContent(systemMsg.content);
if (content) {
systemPrompt = `=== 系统指令 (永远置顶) ===\n${content}\n\n`;
}
}
// 2. 区分历史和当前消息
// 找到最后一条 user 消息的索引
let lastUserIndex = -1;
for (let i = messages.length - 1; i >= 0; i--) {
if (messages[i].role === 'user') {
lastUserIndex = i;
break;
}
}
if (lastUserIndex === -1) {
return parseError(ERROR_CODES.NO_USER_MESSAGES);
}
// 3. 构建历史对话 (不包含 system 和 最后一条 user)
const historyMessages = messages.filter((m, index) => {
return m.role !== 'system' && index < lastUserIndex;
});
if (historyMessages.length > 0) {
historyPrompt += `=== 历史对话 (滑动窗口或摘要) ===\n`;
for (const msg of historyMessages) {
const roleName = msg.role === 'user' ? 'User' : 'AI';
const content = await processContent(msg.content);
historyPrompt += `${roleName}: ${content}\n`;
}
historyPrompt += `\n`;
}
// 4. 构建当前输入
const lastUserMsg = messages[lastUserIndex];
const currentContent = await processContent(lastUserMsg.content);
currentPrompt = `=== 当前输入 ===\nUser: ${currentContent}`;
// 5. 合并最终 Prompt
// 注意:这里我们告诉浏览器上传了 imagePaths 里的所有图片
// 并且在 prompt 文本中已经用 [图片N] 标记了它们的位置
const finalPrompt = systemPrompt + historyPrompt + currentPrompt;
return {
success: true,
data: {
prompt: finalPrompt,
imagePaths, // 包含历史和当前的图片路径,按顺序排列
modelId,
modelName: modelId,
isStreaming
}
};
}
/**
* 解析生图请求 (原有逻辑保持不变)
*/
async function parseImageRequest(messages, tempDir, imageLimit, modelId, isStreaming, getImagePolicy) {
// 筛选用户消息
const userMessages = messages.filter(m => m.role === 'user');
if (userMessages.length === 0) {
return parseError(ERROR_CODES.NO_USER_MESSAGES);
}
const lastMessage = userMessages[userMessages.length - 1];
let prompt = '';
const imagePaths = [];
let imageCount = 0;
// 解析内容
if (Array.isArray(lastMessage.content)) {
for (const item of lastMessage.content) {
if (item.type === 'text') {
prompt += item.text + ' ';
} else if (item.type === 'image_url' && item.image_url?.url) {
imageCount++;
// 图片数量检查
if (imageLimit <= 10) {
if (imageCount > imageLimit) {
return parseError(ERROR_CODES.TOO_MANY_IMAGES, `图片数量超过限制(最大 ${imageLimit} 张)`);
}
} else {
if (imageCount > 10) continue;
}
// 处理 data URL
const url = item.image_url.url;
if (url.startsWith('data:image')) {
const imagePath = await saveBase64Image(url, tempDir);
if (imagePath) {
imagePaths.push(imagePath);
}
}
}
}
} else {
prompt = lastMessage.content;
}
prompt = prompt.trim();
// 图片策略校验
const hasImage = imagePaths.length > 0;
const policy = modelId ? getImagePolicy(modelId) : IMAGE_POLICY.OPTIONAL;
if (policy === IMAGE_POLICY.REQUIRED && !hasImage) {
return parseError(ERROR_CODES.IMAGE_REQUIRED, `模型 ${modelId} 需要参考图`);
}
if (policy === IMAGE_POLICY.FORBIDDEN && hasImage) {
return parseError(ERROR_CODES.IMAGE_FORBIDDEN, `模型 ${modelId} 不支持图片输入`);
}
return {
success: true,
data: {
prompt,
imagePaths,
modelId,
modelName: modelId,
isStreaming
}
};
}
/**
* 保存 Base64 图片到临时文件
*/
async function saveBase64Image(dataUrl, tempDir) {
const matches = dataUrl.match(/^data:([A-Za-z-+\/]+);base64,(.+)$/);
if (!matches || matches.length !== 3) {
return null;
}
try {
const buffer = Buffer.from(matches[2], 'base64');
// 压缩图片
const processedBuffer = await sharp(buffer)
.jpeg({ quality: 90 })
.toBuffer();
const filename = `img_${Date.now()}_${Math.random().toString(36).substring(7)}.jpg`;
const filePath = path.join(tempDir, filename);
fs.writeFileSync(filePath, processedBuffer);
return filePath;
} catch (e) {
return null;
}
}
```
### 配套修改建议
1. **修改 `src/backend/registry.js` (可选,推荐)**:
`models` 数组中添加 `type: 'text'``type: 'image'` 字段,然后在 `getModelsForAdapter` 中暴露出来。这样 `parseChat.js` 就可以通过 `registry.getModelType(id)` 来精准判断,而不是靠猜。
2. **修改适配器代码**:
目前的适配器(如 `gemini_biz.js`, `zai_is.js`)的 `generateImage` 函数接收的是:
```javascript
async function generateImage(context, prompt, imgPaths, modelId, meta = {})
```
对于文本模型,`prompt` 现在将是一个包含了系统指令、历史对话和当前输入的**长字符串**。`imgPaths` 将包含**所有**涉及到的图片。
* 适配器的逻辑通常是:先上传所有 `imgPaths`,然后在输入框填入 `prompt`,最后发送。
* **注意**:你现有的 `gemini_biz.js` 和 `zai_is.js` 适配器逻辑是先 `pasteImages` 然后 `fillPrompt`。
* 这对于新的文本逻辑**正好适用**!因为图片占位符 `[图片1]` 已经在 `prompt` 文本里了,而实际的图片文件通过 `imgPaths` 传给适配器上传,适配器上传后,输入框里会有图片缩略图,加上我们输入的带有 `[图片1]` 的长文本,AI 就能理解这个对应关系(或者至少作为上下文)。
### 效果预期
当客户端发送如下请求时:
```json
{
"model": "gpt-4-text",
"messages": [
{"role": "system", "content": "你是一个猫娘"},
{"role": "user", "content": "你好"},
{"role": "assistant", "content": "喵~"},
{"role": "user", "content": ["看看这个", {"type": "image_url", "url": "..."}]}
]
}
```
服务器会解析为:
**prompt (传递给适配器输入框的文本):**
```text
=== 系统指令 (永远置顶) ===
你是一个猫娘
=== 历史对话 (滑动窗口或摘要) ===
User: 你好
AI: 喵~
=== 当前输入 ===
User: 看看这个[图片1]
```
**imagePaths (传递给适配器上传的文件):**
`['/tmp/img_xxx.jpg']`
这样你就成功地在一个单轮对话的网页接口上,模拟出了带上下文和多模态理解的多轮对话体验。