feat: chat content use tiktoken count

src/service/utils/tools.ts

@@ -2,10 +2,12 @@ import type { NextApiRequest } from 'next';
 import crypto from 'crypto';
 import jwt from 'jsonwebtoken';
 import { ChatItemType } from '@/types/chat';
-import { encode } from 'gpt-token-utils';
 import { OpenApi, User } from '../mongo';
 import { formatPrice } from '@/utils/user';
 import { ERROR_ENUM } from '../errorCode';
+import { countChatTokens } from '@/utils/tools';
+import { ChatCompletionRequestMessageRoleEnum } from 'openai';
+import { ChatModelEnum } from '@/constants/model';
 
 /* 密码加密 */
 export const hashPassword = (psw: string) => {
@@ -86,8 +88,16 @@ export const authOpenApiKey = async (req: NextApiRequest) => {
 export const httpsAgent = (fast: boolean) =>
   fast ? global.httpsAgentFast : global.httpsAgentNormal;
 
-/* tokens 截断 */
-export const openaiChatFilter = (prompts: ChatItemType[], maxTokens: number) => {
+/* 聊天内容 tokens 截断 */
+export const openaiChatFilter = ({
+  model,
+  prompts,
+  maxTokens
+}: {
+  model: `${ChatModelEnum}`;
+  prompts: ChatItemType[];
+  maxTokens: number;
+}) => {
   const formatPrompts = prompts.map((item) => ({
     obj: item.obj,
     value: item.value
@@ -97,41 +107,64 @@ export const openaiChatFilter = (prompts: ChatItemType[], maxTokens: number) =>
       .trim()
   }));
 
-  let res: ChatItemType[] = [];
-
+  let chats: ChatItemType[] = [];
   let systemPrompt: ChatItemType | null = null;
 
   //  System 词保留
   if (formatPrompts[0]?.obj === 'SYSTEM') {
     systemPrompt = formatPrompts.shift() as ChatItemType;
-    maxTokens -= encode(formatPrompts[0].value).length;
   }
 
-  // 从后往前截取
+  // 格式化文本内容成 chatgpt 格式
+  const map = {
+    Human: ChatCompletionRequestMessageRoleEnum.User,
+    AI: ChatCompletionRequestMessageRoleEnum.Assistant,
+    SYSTEM: ChatCompletionRequestMessageRoleEnum.System
+  };
+
+  let messages: { role: ChatCompletionRequestMessageRoleEnum; content: string }[] = [];
+
+  // 从后往前截取对话内容
   for (let i = formatPrompts.length - 1; i >= 0; i--) {
-    const tokens = encode(formatPrompts[i].value).length;
-    res.unshift(formatPrompts[i]);
+    chats.unshift(formatPrompts[i]);
+
+    messages = (systemPrompt ? [systemPrompt, ...chats] : chats).map((item) => ({
+      role: map[item.obj],
+      content: item.value
+    }));
+
+    const tokens = countChatTokens({
+      model,
+      messages
+    });
+
     /* 整体 tokens 超出范围 */
     if (tokens >= maxTokens) {
       break;
     }
-
-    maxTokens -= tokens;
   }
 
-  return systemPrompt ? [systemPrompt, ...res] : res;
+  return messages;
 };
 
 /* system 内容截断 */
-export const systemPromptFilter = (prompts: string[], maxTokens: number) => {
+export const systemPromptFilter = ({
+  model,
+  prompts,
+  maxTokens
+}: {
+  model: 'gpt-4' | 'gpt-4-32k' | 'gpt-3.5-turbo';
+  prompts: string[];
+  maxTokens: number;
+}) => {
   let splitText = '';
 
   // 从前往前截取
   for (let i = 0; i < prompts.length; i++) {
-    const prompt = prompts[i];
+    const prompt = prompts[i].replace(/\n+/g, '\n');
 
     splitText += `${prompt}\n`;
-    const tokens = encode(splitText).length;
+    const tokens = countChatTokens({ model, messages: [{ role: 'system', content: splitText }] });
     if (tokens >= maxTokens) {
       break;
     }