MigrateRepos/glm-free-api
1
0
Fork 0
mirror of https://github.com/LLM-Red-Team/glm-free-api.git synced 2025-01-23 21:31:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

支持GLM-4-Plus以及Zero思考推理模型

Browse Source
This commit is contained in:
Vinlic 2024-12-31 11:32:25 +08:00
parent 57b042d187
commit 719e3b682f
6 changed files with 107 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
README.md
View File

@ -9,7 +9,7 @@
![](https://img.shields.io/github/forks/llm-red-team/glm-free-api.svg)
![](https://img.shields.io/docker/pulls/vinlic/glm-free-api.svg)
支持高速流式输出、支持多轮对话、支持智能体对话、支持视频生成、支持AI绘图、支持联网搜索、支持长文档解读、支持图像解析，零配置部署，多路token支持，自动清理会话痕迹。
支持GLM-4-Plus高速流式输出、支持多轮对话、支持智能体对话、支持Zero思考推理模型、支持视频生成、支持AI绘图、支持联网搜索、支持长文档解读、支持图像解析，零配置部署，多路token支持，自动清理会话痕迹。
与ChatGPT接口完全兼容。
@ -286,8 +286,10 @@ Authorization: Bearer [refresh_token]
请求数据：
```json
{
// 如果使用智能体请填写智能体ID到此处，否则可以乱填
"model": "glm4",
// 默认模型：glm-4-plus
// zero思考推理模型：glm-4-zero / glm-4-think
// 如果使用智能体请填写智能体ID到此处
"model": "glm-4-plus",
// 目前多轮对话基于消息合并实现，某些场景可能导致能力下降且受单轮最大token数限制
// 如果您想获得原生的多轮对话体验，可以传入首轮消息获得的id，来接续上下文
// "conversation_id": "65f6c28546bae1f0fbb532de",
@ -307,7 +309,7 @@ Authorization: Bearer [refresh_token]
{
// 如果想获得原生多轮对话体验，此id，你可以传入到下一轮对话的conversation_id来接续上下文
"id": "65f6c28546bae1f0fbb532de",
"model": "glm4",
"model": "glm-4",
"object": "chat.completion",
"choices": [
{
@ -432,7 +434,7 @@ Authorization: Bearer [refresh_token]
```json
{
// 如果使用智能体请填写智能体ID到此处，否则可以乱填
"model": "glm4",
"model": "glm-4",
"messages": [
{
"role": "user",
@ -459,7 +461,7 @@ Authorization: Bearer [refresh_token]
```json
{
"id": "cnmuo7mcp7f9hjcmihn0",
"model": "glm4",
"model": "glm-4",
"object": "chat.completion",
"choices": [
{

12
README_EN.md
View File

@ -289,8 +289,10 @@ Authorization: Bearer [refresh_token]
Request data:
```json
{
// Except using the Agent to fill the ID, fill in the model name as you like.
"model": "glm4",
// Default model: glm-4-plus
// zero thinking model: glm-4-zero / glm-4-think
// If using the Agent, fill in the Agent ID here
"model": "glm-4",
// Currently, multi-round conversations are realized based on message merging, which in some scenarios may lead to capacity degradation and is limited by the maximum number of tokens in a single round.
// If you want a native multi-round dialog experience, you can pass in the ids obtained from the last round of messages to pick up the context
// "conversation_id": "65f6c28546bae1f0fbb532de",
@ -309,7 +311,7 @@ Response data：
```json
{
"id": "65f6c28546bae1f0fbb532de",
"model": "glm4",
"model": "glm-4",
"object": "chat.completion",
"choices": [
{
@ -434,7 +436,7 @@ Request data:
```json
{
// 如果使用智能体请填写智能体ID到此处，否则可以乱填
"model": "glm4",
"model": "glm-4",
"messages": [
{
"role": "user",
@ -461,7 +463,7 @@ Response data:
```json
{
"id": "cnmuo7mcp7f9hjcmihn0",
"model": "glm4",
"model": "glm-4",
"object": "chat.completion",
"choices": [
{

2
package.json
View File

@ -1,6 +1,6 @@
{
"name": "glm-free-api",
"version": "0.0.33",
"version": "0.0.34",
"description": "GLM Free API Server",
"type": "module",
"main": "dist/index.js",

95
src/api/controllers/chat.ts
View File

@ -17,6 +17,8 @@ import util from "@/lib/util.ts";
const MODEL_NAME = "glm";
// 默认的智能体ID，GLM4
const DEFAULT_ASSISTANT_ID = "65940acff94777010aa6b796";
// zero推理模型智能体ID
const ZERO_ASSISTANT_ID = "676411c38945bbc58a905d31";
// access_token有效期
const ACCESS_TOKEN_EXPIRES = 3600;
// 最大重试次数
@ -165,13 +167,13 @@ async function removeConversation(
*
* @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 model 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,
model = MODEL_NAME,
refConvId = "",
retryCount = 0
) {
@ -189,6 +191,13 @@ async function createCompletion(
// 如果引用对话ID不正确则重置引用
if (!/[0-9a-zA-Z]{24}/.test(refConvId)) refConvId = "";
let assistantId = /^[a-z0-9]{24,}$/.test(model) ? model : undefined;
if(model.indexOf('think') != -1 || model.indexOf('zero') != -1) {
assistantId = ZERO_ASSISTANT_ID;
logger.info('使用思考模型');
}
// 请求流
const token = await acquireToken(refreshToken);
const result = await axios.post(
@ -200,8 +209,11 @@ async function createCompletion(
meta_data: {
channel: "",
draft_id: "",
if_plus_model: true,
input_question_type: "xxxx",
is_test: false,
platform: "pc",
quote_log_id: ""
},
},
{
@ -231,7 +243,7 @@ async function createCompletion(
const streamStartTime = util.timestamp();
// 接收流为输出文本
const answer = await receiveStream(result.data);
const answer = await receiveStream(model, result.data);
logger.success(
`Stream has completed transfer ${util.timestamp() - streamStartTime}ms`
);
@ -251,7 +263,7 @@ async function createCompletion(
return createCompletion(
messages,
refreshToken,
assistantId,
model,
refConvId,
retryCount + 1
);
@ -266,13 +278,13 @@ async function createCompletion(
*
* @param messages 参考gpt系列消息格式，多轮对话请完整提供上下文
* @param refreshToken 用于刷新access_token的refresh_token
* @param assistantId 智能体ID，默认使用GLM4原版
* @param model 智能体ID，默认使用GLM4原版
* @param retryCount 重试次数
*/
async function createCompletionStream(
messages: any[],
refreshToken: string,
assistantId = DEFAULT_ASSISTANT_ID,
model = MODEL_NAME,
refConvId = "",
retryCount = 0
) {
@ -290,6 +302,13 @@ async function createCompletionStream(
// 如果引用对话ID不正确则重置引用
if (!/[0-9a-zA-Z]{24}/.test(refConvId)) refConvId = "";
let assistantId = /^[a-z0-9]{24,}$/.test(model) ? model : undefined;
if(model.indexOf('think') != -1 || model.indexOf('zero') != -1) {
assistantId = ZERO_ASSISTANT_ID;
logger.info('使用思考模型');
}
// 请求流
const token = await acquireToken(refreshToken);
const result = await axios.post(
@ -301,8 +320,11 @@ async function createCompletionStream(
meta_data: {
channel: "",
draft_id: "",
if_plus_model: true,
input_question_type: "xxxx",
is_test: false,
platform: "pc",
quote_log_id: ""
},
},
{
@ -354,7 +376,7 @@ async function createCompletionStream(
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) => {
return createTransStream(model, result.data, (convId: string) => {
logger.success(
`Stream has completed transfer ${util.timestamp() - streamStartTime}ms`
);
@ -372,7 +394,7 @@ async function createCompletionStream(
return createCompletionStream(
messages,
refreshToken,
assistantId,
model,
refConvId,
retryCount + 1
);
@ -407,8 +429,11 @@ async function generateImages(
meta_data: {
channel: "",
draft_id: "",
if_plus_model: true,
input_question_type: "xxxx",
is_test: false,
platform: "pc",
quote_log_id: ""
},
},
{
@ -904,14 +929,15 @@ function checkResult(result: AxiosResponse, refreshToken: string) {
/**
* 从流接收完整的消息内容
*
* @param model 模型
* @param stream 消息流
*/
async function receiveStream(stream: any): Promise<any> {
async function receiveStream(model: string, stream: any): Promise<any> {
return new Promise((resolve, reject) => {
// 消息初始化
const data = {
id: "",
model: MODEL_NAME,
model,
object: "chat.completion",
choices: [
{
@ -923,6 +949,8 @@ async function receiveStream(stream: any): Promise<any> {
usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
created: util.unixTimestamp(),
};
const isSilentModel = model.indexOf('silent') != -1;
let thinkingText = "";
let toolCall = false;
let codeGenerating = false;
let textChunkLength = 0;
@ -930,6 +958,7 @@ async function receiveStream(stream: any): Promise<any> {
let lastExecutionOutput = "";
let textOffset = 0;
let refContent = "";
logger.info(`是否静默模型: ${isSilentModel}`);
const parser = createParser((event) => {
try {
if (event.type !== "event") return;
@ -957,6 +986,7 @@ async function receiveStream(stream: any): Promise<any> {
textChunkLength = 0;
innerStr += "\n";
}
if (type == "text") {
if (toolCall) {
innerStr += "\n";
@ -965,11 +995,20 @@ async function receiveStream(stream: any): Promise<any> {
}
if (partStatus == "finish") textChunkLength = text.length;
return innerStr + text;
} else if (
} else if (type == "text_thinking" && !isSilentModel) {
if (toolCall) {
innerStr += "\n";
textOffset++;
toolCall = false;
}
thinkingText = text;
return innerStr;
}else if (
type == "quote_result" &&
status == "finish" &&
meta_data &&
_.isArray(meta_data.metadata_list)
_.isArray(meta_data.metadata_list) &&
!isSilentModel
) {
refContent = meta_data.metadata_list.reduce((meta, v) => {
return meta + `${v.title} - ${v.url}\n`;
@ -1032,6 +1071,8 @@ async function receiveStream(stream: any): Promise<any> {
);
data.choices[0].message.content += chunk;
} else {
if(thinkingText)
data.choices[0].message.content = `[思考开始]\n${thinkingText}[思考结束]\n\n${data.choices[0].message.content}`;
data.choices[0].message.content =
data.choices[0].message.content.replace(
/【\d+†(来源|源|source)】/g,
@ -1059,18 +1100,22 @@ async function receiveStream(stream: any): Promise<any> {
*
* 将流格式转换为gpt兼容流格式
*
* @param model 模型
* @param stream 消息流
* @param endCallback 传输结束回调
*/
function createTransStream(stream: any, endCallback?: Function) {
function createTransStream(model: string, stream: any, endCallback?: Function) {
// 消息创建时间
const created = util.unixTimestamp();
// 创建转换流
const transStream = new PassThrough();
const isSilentModel = model.indexOf('silent') != -1;
let content = "";
let thinking = false;
let toolCall = false;
let codeGenerating = false;
let textChunkLength = 0;
let thinkingText = "";
let codeTemp = "";
let lastExecutionOutput = "";
let textOffset = 0;
@ -1078,7 +1123,7 @@ function createTransStream(stream: any, endCallback?: Function) {
transStream.write(
`data: ${JSON.stringify({
id: "",
model: MODEL_NAME,
model,
object: "chat.completion.chunk",
choices: [
{
@ -1116,6 +1161,11 @@ function createTransStream(stream: any, endCallback?: Function) {
innerStr += "\n";
}
if (type == "text") {
if(thinking) {
innerStr += "[思考结束]\n\n"
textOffset = thinkingText.length + 8;
thinking = false;
}
if (toolCall) {
innerStr += "\n";
textOffset++;
@ -1123,11 +1173,26 @@ function createTransStream(stream: any, endCallback?: Function) {
}
if (partStatus == "finish") textChunkLength = text.length;
return innerStr + text;
} else if (type == "text_thinking" && !isSilentModel) {
if(!thinking) {
innerStr += "[思考开始]\n";
textOffset = 7;
thinking = true;
}
if (toolCall) {
innerStr += "\n";
textOffset++;
toolCall = false;
}
if (partStatus == "finish") textChunkLength = text.length;
thinkingText += text.substring(thinkingText.length, text.length);
return innerStr + text;
} else if (
type == "quote_result" &&
status == "finish" &&
meta_data &&
_.isArray(meta_data.metadata_list)
_.isArray(meta_data.metadata_list) &&
!isSilentModel
) {
const searchText =
meta_data.metadata_list.reduce(

9
src/api/routes/chat.ts
View File

@ -5,6 +5,9 @@ import Response from '@/lib/response/Response.ts';
import chat from '@/api/controllers/chat.ts';
import logger from '@/lib/logger.ts';
// zero推理模型智能体ID
const ZERO_ASSISTANT_ID = "676411c38945bbc58a905d31";
export default {
prefix: '/v1/chat',
@ -21,15 +24,15 @@ export default {
// 随机挑选一个refresh_token
const token = _.sample(tokens);
const { model, conversation_id: convId, messages, stream } = request.body;
const assistantId = /^[a-z0-9]{24,}$/.test(model) ? model : undefined
if (stream) {
const stream = await chat.createCompletionStream(messages, token, assistantId, convId);
const stream = await chat.createCompletionStream(messages, token, model, convId);
return new Response(stream, {
type: "text/event-stream"
});
}
else
return await chat.createCompletion(messages, token, assistantId, convId);
return await chat.createCompletion(messages, token, model, convId);
}
}

5
src/api/routes/models.ts
View File

@ -18,6 +18,11 @@ export default {
"object": "model",
"owned_by": "glm-free-api"
},
{
"id": "glm-4-plus",
"object": "model",
"owned_by": "glm-free-api"
},
{
"id": "glm-4v",
"object": "model",