首轮不注入注意力prompt

optimize code in messagesPrepare

README.md

@@ -9,9 +9,19 @@
 
 与ChatGPT接口完全兼容。
 
+还有以下四个free-api欢迎关注：
+
+阶跃星辰 (跃问StepChat) 接口转API [step-free-api](https://github.com/LLM-Red-Team/step-free-api)
+
+阿里通义 (Qwen) 接口转API [qwen-free-api](https://github.com/LLM-Red-Team/qwen-free-api)
+
+ZhipuAI (智谱清言) 接口转API [glm-free-api](https://github.com/LLM-Red-Team/glm-free-api)
+
+聆心智能 (Emohaa) 接口转API [emohaa-free-api](https://github.com/LLM-Red-Team/emohaa-free-api)
+
 ## 目录
 
-* [声明](#声明)
+* [免责声明](#免责声明)
 * [在线体验](#在线体验)
 * [效果示例](#效果示例)
 * [接入准备](#接入准备)
@@ -26,13 +36,15 @@
 * [注意事项](#注意事项)
   * [Nginx反代优化](#Nginx反代优化)
 
-## 声明
+## 免责声明
 
-仅限自用，禁止对外提供服务或商用，避免对官方造成服务压力，否则风险自担！
+**本组织和个人不接受任何资金捐助和交易，此项目是纯粹研究交流学习性质！**
 
-仅限自用，禁止对外提供服务或商用，避免对官方造成服务压力，否则风险自担！
+**仅限自用，禁止对外提供服务或商用，避免对官方造成服务压力，否则风险自担！**
 
-仅限自用，禁止对外提供服务或商用，避免对官方造成服务压力，否则风险自担！
+**仅限自用，禁止对外提供服务或商用，避免对官方造成服务压力，否则风险自担！**
+
+**仅限自用，禁止对外提供服务或商用，避免对官方造成服务压力，否则风险自担！**
 
 ## 在线体验
 
@@ -66,10 +78,6 @@ https://udify.app/chat/Po0F6BMJ15q5vu2P
 
 ![响应流畅度一致](https://github.com/LLM-Red-Team/kimi-free-api/assets/20235341/48c7ec00-2b03-46c4-95d0-452d3075219b)
 
-### 100线程并发测试
-
-![100线程并发测试](./doc/example-7.jpg)
-
 ## 接入准备
 
 从 [kimi.moonshot.cn](https://kimi.moonshot.cn) 获取refresh_token
@@ -84,7 +92,7 @@ https://udify.app/chat/Po0F6BMJ15q5vu2P
 
 ### 多账号接入
 
-目前kimi限制普通账号每3小时内只能进行30轮长文本的问答，你可以通过提供多个账号的refresh_token并使用`,`拼接提供：
+目前kimi限制普通账号每3小时内只能进行30轮长文本的问答（短文本不限），你可以通过提供多个账号的refresh_token并使用`,`拼接提供：
 
 `Authorization: Bearer TOKEN1,TOKEN2,TOKEN3`
 
@@ -201,6 +209,8 @@ Authorization: Bearer [refresh_token]
 请求数据：
 ```json
 {
+    // 模型名称随意填写，如果不希望输出检索过程模型名称请包含silent_search
+    "model": "kimi",
     "messages": [
         {
             "role": "user",
@@ -254,6 +264,8 @@ Authorization: Bearer [refresh_token]
 请求数据：
 ```json
 {
+    // 模型名称随意填写，如果不希望输出检索过程模型名称请包含silent_search
+    "model": "kimi",
     "messages": [
         {
             "role": "user",
@@ -318,6 +330,8 @@ Authorization: Bearer [refresh_token]
 请求数据：
 ```json
 {
+    // 模型名称随意填写，如果不希望输出检索过程模型名称请包含silent_search
+    "model": "kimi",
     "messages": [
         {
             "role": "user",
@@ -386,4 +400,8 @@ keepalive_timeout 120;
 
 ### Token统计
 
-由于推理侧不再kimi-free-api，因此token不可统计，将以固定数字返回。
+由于推理侧不在kimi-free-api，因此token不可统计，将以固定数字返回!!!!!
+
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=LLM-Red-Team/kimi-free-api&type=Date)](https://star-history.com/#LLM-Red-Team/kimi-free-api&Date)

package.json

@@ -1,6 +1,6 @@
 {
   "name": "kimi-free-api",
-  "version": "0.0.15",
+  "version": "0.0.22",
   "description": "Kimi Free API Server",
   "type": "module",
   "main": "dist/index.js",

src/api/controllers/chat.ts

@@ -164,12 +164,13 @@ async function removeConversation(convId: string, refreshToken: string) {
 /**
  * 同步对话补全
  * 
+ * @param model 模型名称
  * @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 useSearch 是否开启联网搜索
  * @param retryCount 重试次数
  */
-async function createCompletion(messages: any[], refreshToken: string, useSearch = true, retryCount = 0) {
+async function createCompletion(model = MODEL_NAME, messages: any[], refreshToken: string, useSearch = true, retryCount = 0) {
   return (async () => {
     logger.info(messages);
 
@@ -204,7 +205,7 @@ async function createCompletion(messages: any[], refreshToken: string, useSearch
 
     const streamStartTime = util.timestamp();
     // 接收流为输出文本
-    const answer = await receiveStream(convId, result.data);
+    const answer = await receiveStream(model, convId, result.data);
     logger.success(`Stream has completed transfer ${util.timestamp() - streamStartTime}ms`);
 
     // 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
@@ -219,7 +220,7 @@ async function createCompletion(messages: any[], refreshToken: string, useSearch
         logger.warn(`Try again after ${RETRY_DELAY / 1000}s...`);
         return (async () => {
           await new Promise(resolve => setTimeout(resolve, RETRY_DELAY));
-          return createCompletion(messages, refreshToken, useSearch, retryCount + 1);
+          return createCompletion(model, messages, refreshToken, useSearch, retryCount + 1);
         })();
       }
       throw err;
@@ -229,12 +230,13 @@ async function createCompletion(messages: any[], refreshToken: string, useSearch
 /**
  * 流式对话补全
  * 
+ * @param model 模型名称
  * @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 useSearch 是否开启联网搜索
  * @param retryCount 重试次数
  */
-async function createCompletionStream(messages: any[], refreshToken: string, useSearch = true, retryCount = 0) {
+async function createCompletionStream(model = MODEL_NAME, messages: any[], refreshToken: string, useSearch = true, retryCount = 0) {
   return (async () => {
     logger.info(messages);
 
@@ -268,7 +270,7 @@ async function createCompletionStream(messages: any[], refreshToken: string, use
     });
     const streamStartTime = util.timestamp();
     // 创建转换流将消息格式转换为gpt兼容格式
-    return createTransStream(convId, result.data, () => {
+    return createTransStream(model, convId, result.data, () => {
       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)
@@ -281,7 +283,7 @@ async function createCompletionStream(messages: any[], refreshToken: string, use
         logger.warn(`Try again after ${RETRY_DELAY / 1000}s...`);
         return (async () => {
           await new Promise(resolve => setTimeout(resolve, RETRY_DELAY));
-          return createCompletionStream(messages, refreshToken, useSearch, retryCount + 1);
+          return createCompletionStream(model, messages, refreshToken, useSearch, retryCount + 1);
         })();
       }
       throw err;
@@ -297,7 +299,7 @@ async function createCompletionStream(messages: any[], refreshToken: string, use
  */
 async function fakeRequest(refreshToken: string) {
   const token = await acquireToken(refreshToken);
-  const options =  {
+  const options = {
     headers: {
       Authorization: `Bearer ${token}`,
       Referer: `https://kimi.moonshot.cn/`,
@@ -326,21 +328,27 @@ async function fakeRequest(refreshToken: string) {
  * @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;
-        // kimi-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']);
-      });
-    }
+  const urls = [];
+  // 如果没有消息，则返回[]
+  if (!messages.length) {
     return urls;
-  }, []);
+  }
+  // 只获取最新的消息
+  const lastMessage = messages[messages.length - 1];
+  if (_.isArray(lastMessage.content)) {
+    lastMessage.content.forEach(v => {
+      if (!_.isObject(v) || !['file', 'image_url'].includes(v['type']))
+        return;
+      // kimi-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']);
+    });
+  }
+  logger.info("本次请求上传：" + urls.length + "个文件");
+  return urls;
 }
 
 /**
@@ -354,17 +362,39 @@ function extractRefFileUrls(messages: any[]) {
  * @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[]) {
+  // 注入消息提升注意力
+  let latestMessage = messages[messages.length - 1];
+  let hasFileOrImage = Array.isArray(latestMessage.content)
+    && latestMessage.content.some(v => (typeof v === 'object' && ['file', 'image_url'].includes(v['type'])));
+  // 第二轮开始注入system prompt
+  if (messages.length > 2) {
+    if (hasFileOrImage) {
+      let newFileMessage = {
+        "content": "关注用户最新发送文件和消息",
+        "role": "system"
+      };
+      messages.splice(messages.length - 1, 0, newFileMessage);
+      logger.info("注入提升尾部文件注意力system prompt");
+    } else {
+      let newTextMessage = {
+        "content": "关注用户最新的消息",
+        "role": "system"
+      };
+      messages.splice(messages.length - 1, 0, newTextMessage);
+      logger.info("注入提升尾部消息注意力system prompt");
+    }
+  }
+
   const content = messages.reduce((content, message) => {
-    if (_.isArray(message.content)) {
+    if (Array.isArray(message.content)) {
       return message.content.reduce((_content, v) => {
-        if (!_.isObject(v) || v['type'] != 'text')
-          return _content;
-        return _content + (v['text'] || '');
+        if (!_.isObject(v) || v['type'] != 'text') return _content;
+        return _content + `${message.role || "user"}:${v["text"] || ""}\n`;
       }, content);
     }
     return content += `${message.role || 'user'}:${wrapUrlsToTags(message.content)}\n`;
   }, '');
-
+  logger.info("\n对话合并：\n" + content);
   return [
     { role: 'user', content }
   ]
@@ -472,8 +502,8 @@ async function uploadFile(fileUrl: string, refreshToken: string) {
     data: fileData,
     // 100M限制
     maxBodyLength: FILE_MAX_SIZE,
-    // 60秒超时
-    timeout: 60000,
+    // 120秒超时
+    timeout: 120000,
     headers: {
       'Content-Type': mimeType,
       Authorization: `Bearer ${token}`,
@@ -541,15 +571,16 @@ function checkResult(result: AxiosResponse, refreshToken: string) {
 /**
  * 从流接收完整的消息内容
  * 
+ * @param model 模型名称
  * @param convId 会话ID
  * @param stream 消息流
  */
-async function receiveStream(convId: string, stream: any) {
+async function receiveStream(model: string, convId: string, stream: any) {
   return new Promise((resolve, reject) => {
     // 消息初始化
     const data = {
       id: convId,
-      model: MODEL_NAME,
+      model,
       object: 'chat.completion',
       choices: [
         { index: 0, message: { role: 'assistant', content: '' }, finish_reason: 'stop' }
@@ -558,6 +589,7 @@ async function receiveStream(convId: string, stream: any) {
       created: util.unixTimestamp()
     };
     let refContent = '';
+    const silentSearch = model.indexOf('silent_search') != -1;
     const parser = createParser(event => {
       try {
         if (event.type !== "event") return;
@@ -566,8 +598,9 @@ async function receiveStream(convId: string, stream: any) {
         if (_.isError(result))
           throw new Error(`Stream response invalid: ${event.data}`);
         // 处理消息
-        if (result.event == 'cmpl') {
-          data.choices[0].message.content += result.text;
+        if (result.event == 'cmpl' && result.text) {
+          const exceptCharIndex = result.text.indexOf("<22>");
+          data.choices[0].message.content += result.text.substring(0, exceptCharIndex == -1 ? result.text.length : exceptCharIndex);
         }
         // 处理结束或错误
         else if (result.event == 'all_done' || result.event == 'error') {
@@ -576,7 +609,7 @@ async function receiveStream(convId: string, stream: any) {
           resolve(data);
         }
         // 处理联网搜索
-        else if (result.event == 'search_plus' && result.msg && result.msg.type == 'get_res')
+        else if (!silentSearch && result.event == 'search_plus' && result.msg && result.msg.type == 'get_res')
           refContent += `${result.msg.title}(${result.msg.url})\n`;
         // else
         //   logger.warn(result.event, result);
@@ -598,19 +631,21 @@ async function receiveStream(convId: string, stream: any) {
  * 
  * 将流格式转换为gpt兼容流格式
  * 
+ * @param model 模型名称
  * @param convId 会话ID
  * @param stream 消息流
  * @param endCallback 传输结束回调
  */
-function createTransStream(convId: string, stream: any, endCallback?: Function) {
+function createTransStream(model: string, convId: string, stream: any, endCallback?: Function) {
   // 消息创建时间
   const created = util.unixTimestamp();
   // 创建转换流
   const transStream = new PassThrough();
   let searchFlag = false;
+  const silentSearch = model.indexOf('silent_search') != -1;
   !transStream.closed && transStream.write(`data: ${JSON.stringify({
     id: convId,
-    model: MODEL_NAME,
+    model,
     object: 'chat.completion.chunk',
     choices: [
       { index: 0, delta: { role: 'assistant', content: '' }, finish_reason: null }
@@ -626,12 +661,14 @@ function createTransStream(convId: string, stream: any, endCallback?: Function)
         throw new Error(`Stream response invalid: ${event.data}`);
       // 处理消息
       if (result.event == 'cmpl') {
+        const exceptCharIndex = result.text.indexOf("<22>");
+        const chunk = result.text.substring(0, exceptCharIndex == -1 ? result.text.length : exceptCharIndex);
         const data = `data: ${JSON.stringify({
           id: convId,
-          model: MODEL_NAME,
+          model,
           object: 'chat.completion.chunk',
           choices: [
-            { index: 0, delta: { content: (searchFlag ? '\n' : '') + result.text }, finish_reason: null }
+            { index: 0, delta: { content: (searchFlag ? '\n' : '') + chunk }, finish_reason: null }
           ],
           created
         })}\n\n`;
@@ -643,7 +680,7 @@ function createTransStream(convId: string, stream: any, endCallback?: Function)
       else if (result.event == 'all_done' || result.event == 'error') {
         const data = `data: ${JSON.stringify({
           id: convId,
-          model: MODEL_NAME,
+          model,
           object: 'chat.completion.chunk',
           choices: [
             {
@@ -660,12 +697,12 @@ function createTransStream(convId: string, stream: any, endCallback?: Function)
         endCallback && endCallback();
       }
       // 处理联网搜索
-      else if (result.event == 'search_plus' && result.msg && result.msg.type == 'get_res') {
+      else if (!silentSearch && result.event == 'search_plus' && result.msg && result.msg.type == 'get_res') {
         if (!searchFlag)
           searchFlag = true;
         const data = `data: ${JSON.stringify({
           id: convId,
-          model: MODEL_NAME,
+          model,
           object: 'chat.completion.chunk',
           choices: [
             {

src/api/routes/chat.ts

@@ -19,15 +19,16 @@ export default {
             const tokens = chat.tokenSplit(request.headers.authorization);
             // 随机挑选一个refresh_token
             const token = _.sample(tokens);
+            const model = request.body.model;
             const messages =  request.body.messages;
             if (request.body.stream) {
-                const stream = await chat.createCompletionStream(request.body.messages, token, request.body.use_search);
+                const stream = await chat.createCompletionStream(model, messages, token, request.body.use_search);
                 return new Response(stream, {
                     type: "text/event-stream"
                 });
             }
             else
-                return await chat.createCompletion(messages, token, request.body.use_search);
+                return await chat.createCompletion(model, messages, token, request.body.use_search);
         }
 
     }

src/api/routes/index.ts

@@ -1,5 +1,7 @@
 import chat from "./chat.ts";
+import ping from "./ping.ts";
 
 export default [
-    chat
+    chat,
+    ping
 ];

src/api/routes/ping.ts

@@ -0,0 +1,6 @@
+export default {
+  prefix: '/ping',
+  get: {
+    '': async () => "pong"
+  }
+}