mirror of
https://github.com/labring/FastGPT.git
synced 2025-08-02 20:58:12 +00:00
System optimize (#303)
This commit is contained in:
Archer
37
client/src/utils/common/adapt/message.ts
Normal file
37
client/src/utils/common/adapt/message.ts
Normal file
@@ -0,0 +1,37 @@
|
||||
import type { ChatItemType } from '@/types/chat';
|
||||
import { ChatRoleEnum } from '@/constants/chat';
|
||||
import { ChatCompletionRequestMessageRoleEnum } from 'openai';
|
||||
import type { MessageItemType } from '@/pages/api/openapi/v1/chat/completions';
|
||||
|
||||
const chat2Message = {
|
||||
[ChatRoleEnum.AI]: ChatCompletionRequestMessageRoleEnum.Assistant,
|
||||
[ChatRoleEnum.Human]: ChatCompletionRequestMessageRoleEnum.User,
|
||||
[ChatRoleEnum.System]: ChatCompletionRequestMessageRoleEnum.System
|
||||
};
|
||||
const message2Chat = {
|
||||
[ChatCompletionRequestMessageRoleEnum.System]: ChatRoleEnum.System,
|
||||
[ChatCompletionRequestMessageRoleEnum.User]: ChatRoleEnum.Human,
|
||||
[ChatCompletionRequestMessageRoleEnum.Assistant]: ChatRoleEnum.AI,
|
||||
[ChatCompletionRequestMessageRoleEnum.Function]: 'function'
|
||||
};
|
||||
|
||||
export function adaptRole_Chat2Message(role: `${ChatRoleEnum}`) {
|
||||
return chat2Message[role];
|
||||
}
|
||||
export function adaptRole_Message2Chat(role: `${ChatCompletionRequestMessageRoleEnum}`) {
|
||||
return message2Chat[role];
|
||||
}
|
||||
|
||||
export const adaptChat2GptMessages = ({
|
||||
messages,
|
||||
reserveId
|
||||
}: {
|
||||
messages: ChatItemType[];
|
||||
reserveId: boolean;
|
||||
}): MessageItemType[] => {
|
||||
return messages.map((item) => ({
|
||||
...(reserveId && { dataId: item.dataId }),
|
||||
role: chat2Message[item.obj] || ChatCompletionRequestMessageRoleEnum.System,
|
||||
content: item.value || ''
|
||||
}));
|
||||
};
|
||||
1
client/src/utils/common/tiktoken/cl100k_base.json
Normal file
1
client/src/utils/common/tiktoken/cl100k_base.json
Normal file
File diff suppressed because one or more lines are too long
92
client/src/utils/common/tiktoken/index.ts
Normal file
92
client/src/utils/common/tiktoken/index.ts
Normal file
@@ -0,0 +1,92 @@
|
||||
/* Only the token of gpt-3.5-turbo is used */
|
||||
import { ChatItemType } from '@/types/chat';
|
||||
import { Tiktoken } from 'js-tiktoken/lite';
|
||||
import { adaptChat2GptMessages } from '../adapt/message';
|
||||
import { ChatCompletionRequestMessageRoleEnum } from 'openai';
|
||||
import encodingJson from './cl100k_base.json';
|
||||
|
||||
/* init tikToken obj */
|
||||
export function getTikTokenEnc() {
|
||||
if (typeof window !== 'undefined' && window.TikToken) {
|
||||
return window.TikToken;
|
||||
}
|
||||
if (typeof global !== 'undefined' && global.TikToken) {
|
||||
return global.TikToken;
|
||||
}
|
||||
|
||||
const enc = new Tiktoken(encodingJson);
|
||||
|
||||
if (typeof window !== 'undefined') {
|
||||
window.TikToken = enc;
|
||||
}
|
||||
if (typeof global !== 'undefined') {
|
||||
global.TikToken = enc;
|
||||
}
|
||||
|
||||
return enc;
|
||||
}
|
||||
|
||||
/* count one prompt tokens */
|
||||
export function countPromptTokens(prompt = '', role: `${ChatCompletionRequestMessageRoleEnum}`) {
|
||||
const enc = getTikTokenEnc();
|
||||
const text = `${role}\n${prompt}`;
|
||||
try {
|
||||
const encodeText = enc.encode(text);
|
||||
return encodeText.length + 3; // 补充 role 估算值
|
||||
} catch (error) {
|
||||
return text.length;
|
||||
}
|
||||
}
|
||||
|
||||
/* count messages tokens */
|
||||
export function countMessagesTokens({ messages }: { messages: ChatItemType[] }) {
|
||||
const adaptMessages = adaptChat2GptMessages({ messages, reserveId: true });
|
||||
|
||||
let totalTokens = 0;
|
||||
for (let i = 0; i < adaptMessages.length; i++) {
|
||||
const item = adaptMessages[i];
|
||||
const tokens = countPromptTokens(item.content, item.role);
|
||||
totalTokens += tokens;
|
||||
}
|
||||
|
||||
return totalTokens;
|
||||
}
|
||||
|
||||
export function sliceTextByTokens({ text, length }: { text: string; length: number }) {
|
||||
const enc = getTikTokenEnc();
|
||||
|
||||
try {
|
||||
const encodeText = enc.encode(text);
|
||||
return enc.decode(encodeText.slice(0, length));
|
||||
} catch (error) {
|
||||
return text.slice(0, length);
|
||||
}
|
||||
}
|
||||
|
||||
/* slice messages from top to bottom by maxTokens */
|
||||
export function sliceMessagesTB({
|
||||
messages,
|
||||
maxTokens
|
||||
}: {
|
||||
messages: ChatItemType[];
|
||||
maxTokens: number;
|
||||
}) {
|
||||
const adaptMessages = adaptChat2GptMessages({ messages, reserveId: true });
|
||||
let reduceTokens = maxTokens;
|
||||
let result: ChatItemType[] = [];
|
||||
|
||||
for (let i = 0; i < adaptMessages.length; i++) {
|
||||
const item = adaptMessages[i];
|
||||
|
||||
const tokens = countPromptTokens(item.content, item.role);
|
||||
reduceTokens -= tokens;
|
||||
|
||||
if (tokens > 0) {
|
||||
result.push(messages[i]);
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
return result.length === 0 && messages[0] ? [messages[0]] : result;
|
||||
}
|
||||
@@ -1,8 +1,8 @@
|
||||
import mammoth from 'mammoth';
|
||||
import Papa from 'papaparse';
|
||||
import { getOpenAiEncMap } from './plugin/openai';
|
||||
import { getErrText } from './tools';
|
||||
import { uploadImg, postUploadFiles } from '@/api/support/file';
|
||||
import { countPromptTokens } from './common/tiktoken';
|
||||
|
||||
/**
|
||||
* upload file to mongo gridfs
|
||||
@@ -206,16 +206,7 @@ export const splitText2Chunks = ({ text, maxLen }: { text: string; maxLen: numbe
|
||||
chunks.push(chunk);
|
||||
}
|
||||
|
||||
const tokens = (() => {
|
||||
try {
|
||||
const enc = getOpenAiEncMap();
|
||||
const encodeText = enc.encode(chunks.join(''));
|
||||
const tokens = encodeText.length;
|
||||
return tokens;
|
||||
} catch (error) {
|
||||
return chunks.join('').length;
|
||||
}
|
||||
})();
|
||||
const tokens = chunks.reduce((sum, chunk) => sum + countPromptTokens(chunk, 'system'), 0);
|
||||
|
||||
return {
|
||||
chunks,
|
||||
|
||||
@@ -1,8 +0,0 @@
|
||||
import { countOpenAIToken, openAiSliceTextByToken } from './openai';
|
||||
import { gpt_chatItemTokenSlice } from '@/pages/api/openapi/text/gptMessagesSlice';
|
||||
|
||||
export const modelToolMap = {
|
||||
countTokens: countOpenAIToken,
|
||||
sliceText: openAiSliceTextByToken,
|
||||
tokenSlice: gpt_chatItemTokenSlice
|
||||
};
|
||||
@@ -1,100 +0,0 @@
|
||||
import { encoding_for_model } from '@dqbd/tiktoken';
|
||||
import type { ChatItemType } from '@/types/chat';
|
||||
import { ChatRoleEnum } from '@/constants/chat';
|
||||
import { ChatCompletionRequestMessageRoleEnum } from 'openai';
|
||||
import axios from 'axios';
|
||||
import type { MessageItemType } from '@/pages/api/openapi/v1/chat/completions';
|
||||
|
||||
export const getOpenAiEncMap = () => {
|
||||
if (typeof window !== 'undefined' && window.OpenAiEncMap) {
|
||||
return window.OpenAiEncMap;
|
||||
}
|
||||
if (typeof global !== 'undefined' && global.OpenAiEncMap) {
|
||||
return global.OpenAiEncMap;
|
||||
}
|
||||
const enc = encoding_for_model('gpt-3.5-turbo', {
|
||||
'<|im_start|>': 100264,
|
||||
'<|im_end|>': 100265,
|
||||
'<|im_sep|>': 100266
|
||||
});
|
||||
|
||||
if (typeof window !== 'undefined') {
|
||||
window.OpenAiEncMap = enc;
|
||||
}
|
||||
if (typeof global !== 'undefined') {
|
||||
global.OpenAiEncMap = enc;
|
||||
}
|
||||
|
||||
return enc;
|
||||
};
|
||||
|
||||
export const adaptChatItem_openAI = ({
|
||||
messages,
|
||||
reserveId
|
||||
}: {
|
||||
messages: ChatItemType[];
|
||||
reserveId: boolean;
|
||||
}): MessageItemType[] => {
|
||||
const map = {
|
||||
[ChatRoleEnum.AI]: ChatCompletionRequestMessageRoleEnum.Assistant,
|
||||
[ChatRoleEnum.Human]: ChatCompletionRequestMessageRoleEnum.User,
|
||||
[ChatRoleEnum.System]: ChatCompletionRequestMessageRoleEnum.System
|
||||
};
|
||||
return messages.map((item) => ({
|
||||
...(reserveId && { dataId: item.dataId }),
|
||||
role: map[item.obj] || ChatCompletionRequestMessageRoleEnum.System,
|
||||
content: item.value || ''
|
||||
}));
|
||||
};
|
||||
|
||||
export function countOpenAIToken({ messages }: { messages: ChatItemType[] }) {
|
||||
const adaptMessages = adaptChatItem_openAI({ messages, reserveId: true });
|
||||
const token = adaptMessages.reduce((sum, item) => {
|
||||
const text = `${item.role}\n${item.content}`;
|
||||
|
||||
/* use textLen as tokens if encode error */
|
||||
const tokens = (() => {
|
||||
try {
|
||||
const enc = getOpenAiEncMap();
|
||||
const encodeText = enc.encode(text);
|
||||
return encodeText.length + 3; // 补充估算值
|
||||
} catch (error) {
|
||||
return text.length;
|
||||
}
|
||||
})();
|
||||
|
||||
return sum + tokens;
|
||||
}, 0);
|
||||
|
||||
return token;
|
||||
}
|
||||
|
||||
export const openAiSliceTextByToken = ({ text, length }: { text: string; length: number }) => {
|
||||
const enc = getOpenAiEncMap();
|
||||
|
||||
try {
|
||||
const encodeText = enc.encode(text);
|
||||
const decoder = new TextDecoder();
|
||||
return decoder.decode(enc.decode(encodeText.slice(0, length)));
|
||||
} catch (error) {
|
||||
return text.slice(0, length);
|
||||
}
|
||||
};
|
||||
|
||||
export const authOpenAiKey = async (key: string) => {
|
||||
return axios
|
||||
.get('https://ccdbwscohpmu.cloud.sealos.io/openai/v1/dashboard/billing/subscription', {
|
||||
headers: {
|
||||
Authorization: `Bearer ${key}`
|
||||
}
|
||||
})
|
||||
.then((res) => {
|
||||
if (!res.data.access_until) {
|
||||
return Promise.resolve('OpenAI Key 可能无效');
|
||||
}
|
||||
})
|
||||
.catch((err) => {
|
||||
console.log(err);
|
||||
return Promise.reject(err?.response?.data?.error?.message || 'OpenAI Key 可能无效');
|
||||
});
|
||||
};
|
||||
Reference in New Issue
Block a user