feat: maxToken setting

client/src/service/models/model.ts

@@ -47,6 +47,11 @@ const ModelSchema = new Schema({
       type: String,
       default: ''
     },
+    maxToken: {
+      type: Number,
+      default: 4000,
+      min: 100
+    },
     temperature: {
       type: Number,
       min: 0,

client/src/service/utils/chat/index.ts

@@ -12,6 +12,7 @@ import { textAdaptGptResponse } from '@/utils/adapt';
 export type ChatCompletionType = {
   apiKey: string;
   temperature: number;
+  maxToken?: number;
   messages: ChatItemType[];
   chatId?: string;
   [key: string]: any;

client/src/service/utils/chat/openai.ts

@@ -19,22 +19,31 @@ export const chatResponse = async ({
   model,
   apiKey,
   temperature,
+  maxToken = 4000,
   messages,
   stream
 }: ChatCompletionType & { model: `${OpenAiChatEnum}` }) => {
+  const modelTokenLimit = ChatModelMap[model]?.contextMaxToken || 4000;
   const filterMessages = ChatContextFilter({
     model,
     prompts: messages,
-    maxTokens: Math.ceil(ChatModelMap[model].contextMaxToken * 0.85)
+    maxTokens: Math.ceil(modelTokenLimit - 300) // filter token. not response maxToken
   });
 
   const adaptMessages = adaptChatItem_openAI({ messages: filterMessages, reserveId: false });
   const chatAPI = getOpenAIApi();
 
+  const promptsToken = modelToolMap[model].countTokens({
+    messages: filterMessages
+  });
+
+  maxToken = maxToken + promptsToken > modelTokenLimit ? modelTokenLimit - promptsToken : maxToken;
+
   const response = await chatAPI.createChatCompletion(
     {
       model,
       temperature: Number(temperature || 0),
+      max_tokens: maxToken,
       messages: adaptMessages,
       frequency_penalty: 0.5, // 越大，重复内容越少
       presence_penalty: -0.5, // 越大，越容易出现新内容