import { PassThrough } from "stream"; import path from 'path'; import _ from 'lodash'; import mime from 'mime'; import FormData from "form-data"; import axios, { AxiosResponse } from 'axios'; import APIException from "@/lib/exceptions/APIException.ts"; import EX from "@/api/consts/exceptions.ts"; import { createParser } from 'eventsource-parser' import logger from '@/lib/logger.ts'; import util from '@/lib/util.ts'; // 模型名称 const MODEL_NAME = 'glm'; // 默认的智能体ID,GLM4 const DEFAULT_ASSISTANT_ID = '65940acff94777010aa6b796'; // access_token有效期 const ACCESS_TOKEN_EXPIRES = 3600; // 最大重试次数 const MAX_RETRY_COUNT = 3; // 重试延迟 const RETRY_DELAY = 5000; // 伪装headers const FAKE_HEADERS = { 'Accept': '*/*', 'App-Name': 'chatglm', 'Platform': 'pc', 'Origin': 'https://chatglm.cn', 'Sec-Ch-Ua': '"Chromium";v="122", "Not(A:Brand";v="24", "Google Chrome";v="122"', 'Sec-Ch-Ua-Mobile': '?0', 'Sec-Ch-Ua-Platform': '"Windows"', 'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/122.0.0.0 Safari/537.36', 'Version': '0.0.1' }; // 文件最大大小 const FILE_MAX_SIZE = 100 * 1024 * 1024; // access_token映射 const accessTokenMap = new Map(); // access_token请求队列映射 const accessTokenRequestQueueMap: Record = {}; /** * 请求access_token * * 使用refresh_token去刷新获得access_token * * @param refreshToken 用于刷新access_token的refresh_token */ async function requestToken(refreshToken: string) { if (accessTokenRequestQueueMap[refreshToken]) return new Promise(resolve => accessTokenRequestQueueMap[refreshToken].push(resolve)); accessTokenRequestQueueMap[refreshToken] = []; logger.info(`Refresh token: ${refreshToken}`); const result = await (async () => { const result = await axios.post('https://chatglm.cn/chatglm/backend-api/v1/user/refresh', {}, { headers: { Authorization: `Bearer ${refreshToken}`, Referer: 'https://chatglm.cn/main/alltoolsdetail', 'X-Device-Id': util.uuid(false), 'X-Request-Id': util.uuid(false), ...FAKE_HEADERS }, timeout: 15000, validateStatus: () => true }); const { result: _result } = checkResult(result, refreshToken); const { accessToken } = _result; return { accessToken, refreshToken, refreshTime: util.unixTimestamp() + ACCESS_TOKEN_EXPIRES } })() .then(result => { if(accessTokenRequestQueueMap[refreshToken]) { accessTokenRequestQueueMap[refreshToken].forEach(resolve => resolve(result)); delete accessTokenRequestQueueMap[refreshToken]; } logger.success(`Refresh successful`); return result; }) .catch(err => { if(accessTokenRequestQueueMap[refreshToken]) { accessTokenRequestQueueMap[refreshToken].forEach(resolve => resolve(err)); delete accessTokenRequestQueueMap[refreshToken]; } return err; }); if(_.isError(result)) throw result; return result; } /** * 获取缓存中的access_token * * 避免短时间大量刷新token,未加锁,如果有并发要求还需加锁 * * @param refreshToken 用于刷新access_token的refresh_token */ async function acquireToken(refreshToken: string): Promise { let result = accessTokenMap.get(refreshToken); if (!result) { result = await requestToken(refreshToken); accessTokenMap.set(refreshToken, result); } if (util.unixTimestamp() > result.refreshTime) { result = await requestToken(refreshToken); accessTokenMap.set(refreshToken, result); } return result.accessToken; } /** * 移除会话 * * 在对话流传输完毕后移除会话,避免创建的会话出现在用户的对话列表中 * * @param refreshToken 用于刷新access_token的refresh_token */ async function removeConversation(convId: string, refreshToken: string, assistantId = DEFAULT_ASSISTANT_ID) { const token = await acquireToken(refreshToken); const result = await axios.post('https://chatglm.cn/chatglm/backend-api/assistant/conversation/delete', { assistant_id: assistantId, conversation_id: convId }, { headers: { Authorization: `Bearer ${token}`, Referer: `https://chatglm.cn/main/alltoolsdetail`, 'X-Device-Id': util.uuid(false), 'X-Request-Id': util.uuid(false), ...FAKE_HEADERS }, timeout: 15000, validateStatus: () => true }); checkResult(result, refreshToken); } /** * 同步对话补全 * * @param messages 参考gpt系列消息格式,多轮对话请完整提供上下文 * @param refreshToken 用于刷新access_token的refresh_token * @param assistantId 智能体ID,默认使用GLM4原版 * @param retryCount 重试次数 */ async function createCompletion(messages: any[], refreshToken: string, assistantId = DEFAULT_ASSISTANT_ID, retryCount = 0) { return (async () => { logger.info(messages); // 提取引用文件URL并上传获得引用的文件ID列表 const refFileUrls = extractRefFileUrls(messages); const refs = refFileUrls.length ? await Promise.all(refFileUrls.map(fileUrl => uploadFile(fileUrl, refreshToken))) : []; // 请求流 const token = await acquireToken(refreshToken); const result = await axios.post('https://chatglm.cn/chatglm/backend-api/assistant/stream', { assistant_id: assistantId, conversation_id: '', messages: messagesPrepare(messages, refs), meta_data: { channel: '', draft_id: '', input_question_type: 'xxxx', is_test: false } }, { headers: { Authorization: `Bearer ${token}`, Referer: assistantId == DEFAULT_ASSISTANT_ID ? 'https://chatglm.cn/main/alltoolsdetail' : `https://chatglm.cn/main/gdetail/${assistantId}`, 'X-Device-Id': util.uuid(false), 'X-Request-Id': util.uuid(false), ...FAKE_HEADERS }, // 120秒超时 timeout: 120000, validateStatus: () => true, responseType: 'stream' }); if(result.headers['content-type'].indexOf('text/event-stream') == -1) throw new APIException(EX.API_REQUEST_FAILED, `Stream response Content-Type invalid: ${result.headers['content-type']}`); const streamStartTime = util.timestamp(); // 接收流为输出文本 const answer = await receiveStream(result.data); logger.success(`Stream has completed transfer ${util.timestamp() - streamStartTime}ms`); // 异步移除会话 removeConversation(answer.id, refreshToken, assistantId) .catch(err => console.error(err)); return answer; })() .catch(err => { if(retryCount < MAX_RETRY_COUNT) { logger.error(`Stream response error: ${err.stack}`); logger.warn(`Try again after ${RETRY_DELAY / 1000}s...`); return (async () => { await new Promise(resolve => setTimeout(resolve, RETRY_DELAY)); return createCompletion(messages, refreshToken, assistantId, retryCount + 1); })(); } throw err; }); } /** * 流式对话补全 * * @param messages 参考gpt系列消息格式,多轮对话请完整提供上下文 * @param refreshToken 用于刷新access_token的refresh_token * @param assistantId 智能体ID,默认使用GLM4原版 * @param retryCount 重试次数 */ async function createCompletionStream(messages: any[], refreshToken: string, assistantId = DEFAULT_ASSISTANT_ID, retryCount = 0) { return (async () => { logger.info(messages); // 提取引用文件URL并上传获得引用的文件ID列表 const refFileUrls = extractRefFileUrls(messages); const refs = refFileUrls.length ? await Promise.all(refFileUrls.map(fileUrl => uploadFile(fileUrl, refreshToken))) : []; // 请求流 const token = await acquireToken(refreshToken); const result = await axios.post(`https://chatglm.cn/chatglm/backend-api/assistant/stream`, { assistant_id: assistantId, conversation_id: '', messages: messagesPrepare(messages, refs), meta_data: { channel: '', draft_id: '', input_question_type: 'xxxx', is_test: false } }, { headers: { Authorization: `Bearer ${token}`, Referer: assistantId == DEFAULT_ASSISTANT_ID ? 'https://chatglm.cn/main/alltoolsdetail' : `https://chatglm.cn/main/gdetail/${assistantId}`, 'X-Device-Id': util.uuid(false), 'X-Request-Id': util.uuid(false), ...FAKE_HEADERS }, // 120秒超时 timeout: 120000, validateStatus: () => true, responseType: 'stream' }); if(result.headers['content-type'].indexOf('text/event-stream') == -1) { logger.error(`Invalid response Content-Type:`, result.headers['content-type']); const transStream = new PassThrough(); transStream.end(`data: ${JSON.stringify({ id: '', model: MODEL_NAME, object: 'chat.completion.chunk', choices: [ { index: 0, delta: { role: 'assistant', content: '服务暂时不可用,第三方响应错误' }, finish_reason: 'stop' } ], usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 }, created: util.unixTimestamp() })}\n\n`); return transStream; } const streamStartTime = util.timestamp(); // 创建转换流将消息格式转换为gpt兼容格式 return createTransStream(result.data, (convId: string) => { logger.success(`Stream has completed transfer ${util.timestamp() - streamStartTime}ms`); // 流传输结束后异步移除会话,如果消息不合规,此操作可能会抛出数据库错误异常,请忽略 removeConversation(convId, refreshToken, assistantId) .catch(err => console.error(err)); }); })() .catch(err => { if(retryCount < MAX_RETRY_COUNT) { logger.error(`Stream response error: ${err.stack}`); logger.warn(`Try again after ${RETRY_DELAY / 1000}s...`); return (async () => { await new Promise(resolve => setTimeout(resolve, RETRY_DELAY)); return createCompletionStream(messages, refreshToken, assistantId, retryCount + 1); })(); } throw err; }); } /** * 提取消息中引用的文件URL * * @param messages 参考gpt系列消息格式,多轮对话请完整提供上下文 */ function extractRefFileUrls(messages: any[]) { return messages.reduce((urls, message) => { if (_.isArray(message.content)) { message.content.forEach(v => { if (!_.isObject(v) || !['file', 'image_url'].includes(v['type'])) return; // glm-free-api支持格式 if (v['type'] == 'file' && _.isObject(v['file_url']) && _.isString(v['file_url']['url'])) urls.push(v['file_url']['url']); // 兼容gpt-4-vision-preview API格式 else if (v['type'] == 'image_url' && _.isObject(v['image_url']) && _.isString(v['image_url']['url'])) urls.push(v['image_url']['url']); }); } return urls; }, []); } /** * 消息预处理 * * 由于接口只取第一条消息,此处会将多条消息合并为一条,实现多轮对话效果 * 使用”你“这个角色回复”我“这个角色,以第一人称对话\n * 我:旧消息1 * 你:旧消息2 * 我:新消息 * * @param messages 参考gpt系列消息格式,多轮对话请完整提供上下文 */ function messagesPrepare(messages: any[], refs: any[]) { const content = messages.reduce((content, message) => { if (_.isArray(message.content)) { return message.content.reduce((_content, v) => { if (!_.isObject(v) || v['type'] != 'text') return _content; return _content + (v['text'] || ''); }, content) + '\n'; } return content += `${message.role.replace('sytstem', '<|sytstem|>').replace('assistant', '<|assistant|>').replace('user', '<|user|>')}\n${message.content}\n`; }, '') + '<|assistant|>\n'; const fileRefs = refs.filter(ref => !ref.width && !ref.height); const imageRefs = refs.filter(ref => ref.width || ref.height).map(ref => { ref.image_url = ref.file_url; return ref; }); return [ { role: 'user', content: [ { type: 'text', text: content.replace(/\!\[.+\]\(.+\)/g, '') }, ...(fileRefs.length == 0 ? [] : [{ type: 'file', file: fileRefs }]), ...(imageRefs.length == 0 ? [] : [{ type: 'image', image: imageRefs }]) ] } ] } /** * 预检查文件URL有效性 * * @param fileUrl 文件URL */ async function checkFileUrl(fileUrl: string) { if (util.isBASE64Data(fileUrl)) return; const result = await axios.head(fileUrl, { timeout: 15000, validateStatus: () => true }); if (result.status >= 400) throw new APIException(EX.API_FILE_URL_INVALID, `File ${fileUrl} is not valid: [${result.status}] ${result.statusText}`); // 检查文件大小 if (result.headers && result.headers['content-length']) { const fileSize = parseInt(result.headers['content-length'], 10); if (fileSize > FILE_MAX_SIZE) throw new APIException(EX.API_FILE_EXECEEDS_SIZE, `File ${fileUrl} is not valid`); } } /** * 上传文件 * * @param fileUrl 文件URL * @param refreshToken 用于刷新access_token的refresh_token */ async function uploadFile(fileUrl: string, refreshToken: string) { // 预检查远程文件URL可用性 await checkFileUrl(fileUrl); let filename, fileData, mimeType; // 如果是BASE64数据则直接转换为Buffer if (util.isBASE64Data(fileUrl)) { mimeType = util.extractBASE64DataFormat(fileUrl); const ext = mime.getExtension(mimeType); filename = `${util.uuid()}.${ext}`; fileData = Buffer.from(util.removeBASE64DataHeader(fileUrl), 'base64'); } // 下载文件到内存,如果您的服务器内存很小,建议考虑改造为流直传到下一个接口上,避免停留占用内存 else { filename = path.basename(fileUrl); ({ data: fileData } = await axios.get(fileUrl, { responseType: 'arraybuffer', // 100M限制 maxContentLength: FILE_MAX_SIZE, // 60秒超时 timeout: 60000 })); } // 获取文件的MIME类型 mimeType = mimeType || mime.getType(filename); const formData = new FormData(); formData.append('file', fileData, { filename, contentType: mimeType }); // 上传文件到目标OSS const token = await acquireToken(refreshToken); let result = await axios.request({ method: 'POST', url: 'https://chatglm.cn/chatglm/backend-api/assistant/file_upload', data: formData, // 100M限制 maxBodyLength: FILE_MAX_SIZE, // 60秒超时 timeout: 60000, headers: { Authorization: `Bearer ${token}`, Referer: `https://chatglm.cn/`, ...FAKE_HEADERS, ...formData.getHeaders() }, validateStatus: () => true }); const { result: uploadResult } = checkResult(result, refreshToken); return uploadResult; } /** * 检查请求结果 * * @param result 结果 */ function checkResult(result: AxiosResponse, refreshToken: string) { if (!result.data) return null; const { code, status, message } = result.data; if (!_.isFinite(code) && !_.isFinite(status)) return result.data; if (code === 0 || status === 0) return result.data; if (code == 401) accessTokenMap.delete(refreshToken); throw new APIException(EX.API_REQUEST_FAILED, `[请求glm失败]: ${message}`); } /** * 从流接收完整的消息内容 * * @param stream 消息流 */ async function receiveStream(stream: any): Promise { return new Promise((resolve, reject) => { // 消息初始化 const data = { id: '', model: MODEL_NAME, object: 'chat.completion', choices: [ { index: 0, message: { role: 'assistant', content: '' }, finish_reason: 'stop' } ], usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 }, created: util.unixTimestamp() }; let toolCall = false; let textOffset = 0; const parser = createParser(event => { try { if (event.type !== "event") return; // 解析JSON const result = _.attempt(() => JSON.parse(event.data)); if (_.isError(result)) throw new Error(`Stream response invalid: ${event.data}`); if(!data.id && result.conversation_id) data.id = result.conversation_id; if(result.status != 'finish') { const text = result.parts.reduce((str, part) => { const { status, content, meta_data } = part; if(!_.isArray(content)) return str; const partText = content.reduce((innerStr, value) => { const { type, text, image } = value; if(type == 'text') { if(toolCall) { innerStr += '\n'; textOffset++; toolCall = false; } return innerStr + text; } else if(type == 'quote_result' && status == 'finish' && meta_data && _.isArray(meta_data.metadata_list)) { const searchText = meta_data.metadata_list.reduce((meta, v) => meta + `检索 ${v.title}(${v.url}) ...`, '') + '\n'; textOffset += searchText.length; toolCall = true; return innerStr + searchText; } else if(type == 'image' && _.isArray(image) && status == 'finish') { const imageText = image.reduce((imgs, v) => imgs + (/^(http|https):\/\//.test(v.image_url) ? `![图像](${v.image_url || ''})` : ''), '') + '\n'; textOffset += imageText.length; toolCall = true; return innerStr + imageText; } return innerStr; }, ''); return str + partText; }, ''); const chunk = text.substring(data.choices[0].message.content.length - textOffset, text.length); data.choices[0].message.content += chunk; } else { data.choices[0].message.content = data.choices[0].message.content.replace(/【\d+†source】/g, ''); resolve(data); } } catch (err) { logger.error(err); reject(err); } }); // 将流数据喂给SSE转换器 stream.on("data", buffer => parser.feed(buffer.toString())); stream.once("error", err => reject(err)); stream.once("close", () => resolve(data)); }); } /** * 创建转换流 * * 将流格式转换为gpt兼容流格式 * * @param stream 消息流 * @param endCallback 传输结束回调 */ function createTransStream(stream: any, endCallback?: Function) { // 消息创建时间 const created = util.unixTimestamp(); // 创建转换流 const transStream = new PassThrough(); let content = ''; let toolCall = false; let textOffset = 0; !transStream.closed && transStream.write(`data: ${JSON.stringify({ id: '', model: MODEL_NAME, object: 'chat.completion.chunk', choices: [ { index: 0, delta: { role: 'assistant', content: '' }, finish_reason: null } ], created })}\n\n`); const parser = createParser(event => { try { if (event.type !== "event") return; // 解析JSON const result = _.attempt(() => JSON.parse(event.data)); if (_.isError(result)) throw new Error(`Stream response invalid: ${event.data}`); if(result.status != 'finish' && result.status != 'intervene') { const text = result.parts.reduce((str, part) => { const { status, content, meta_data } = part; if(!_.isArray(content)) return str; const partText = content.reduce((innerStr, value) => { const { type, text, image } = value; if(type == 'text') { if(toolCall) { innerStr += '\n'; textOffset++; toolCall = false; } return innerStr + text.replace(/【\d+†source】/g, ''); } else if(type == 'quote_result' && status == 'finish' && meta_data && _.isArray(meta_data.metadata_list)) { const searchText = meta_data.metadata_list.reduce((meta, v) => meta + `检索 ${v.title}(${v.url}) ...`, '') + '\n'; textOffset += searchText.length; toolCall = true; return innerStr + searchText; } else if(type == 'image' && _.isArray(image) && status == 'finish') { const imageText = image.reduce((imgs, v) => imgs + (/^(http|https):\/\//.test(v.image_url) ? `![图像](${v.image_url || ''})` : ''), '') + '\n'; textOffset += imageText.length; toolCall = true; return innerStr + imageText; } return innerStr; }, ''); return str + partText; }, ''); const chunk = text.substring(content.length - textOffset, text.length); if(chunk) { content += chunk; const data = `data: ${JSON.stringify({ id: result.conversation_id, model: MODEL_NAME, object: 'chat.completion.chunk', choices: [ { index: 0, delta: { content: chunk }, finish_reason: null } ], created })}\n\n`; !transStream.closed && transStream.write(data); } } else { const data = `data: ${JSON.stringify({ id: result.conversation_id, model: MODEL_NAME, object: 'chat.completion.chunk', choices: [ { index: 0, delta: result.status == 'intervene' && result.last_error && result.last_error.intervene_text ? { content: `\n\n${result.last_error.intervene_text}` } : {}, finish_reason: 'stop' } ], usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 }, created })}\n\n`; !transStream.closed && transStream.write(data); !transStream.closed && transStream.end('data: [DONE]\n\n'); content = ''; endCallback && endCallback(); } } catch (err) { logger.error(err); !transStream.closed && transStream.end('\n\n'); } }); // 将流数据喂给SSE转换器 stream.on("data", buffer => parser.feed(buffer.toString())); stream.once("error", () => !transStream.closed && transStream.end('data: [DONE]\n\n')); stream.once("close", () => !transStream.closed && transStream.end('data: [DONE]\n\n')); return transStream; } /** * Token切分 * * @param authorization 认证字符串 */ function tokenSplit(authorization: string) { return authorization.replace('Bearer ', '').split(','); } /** * 备用生成cookie * * 暂时还不需要 * * @param refreshToken * @param token */ function generateCookie(refreshToken: string, token: string) { const timestamp = util.unixTimestamp(); const gsTimestamp = timestamp - Math.round(Math.random() * 2592000); return { chatglm_refresh_token: refreshToken, // chatglm_user_id: '', _ga_PMD05MS2V9: `GS1.1.${gsTimestamp}.18.0.${gsTimestamp}.0.0.0`, chatglm_token: token, chatglm_token_expires: util.getDateString('yyyy-MM-dd HH:mm:ss'), abtestid: 'a', // acw_tc: '' } } export default { createCompletion, createCompletionStream, tokenSplit };