mirror of
https://github.com/labring/FastGPT.git
synced 2026-03-05 01:13:16 +08:00
perf: chat framwork
This commit is contained in:
archer
39
src/utils/chat/index.ts
Normal file
39
src/utils/chat/index.ts
Normal file
@@ -0,0 +1,39 @@
|
||||
import { OpenAiChatEnum } from '@/constants/model';
|
||||
import type { ChatModelType } from '@/constants/model';
|
||||
import type { ChatItemSimpleType } from '@/types/chat';
|
||||
import { countOpenAIToken, getOpenAiEncMap, adaptChatItem_openAI } from './openai';
|
||||
|
||||
export type CountTokenType = { messages: ChatItemSimpleType[] };
|
||||
|
||||
export const modelToolMap = {
|
||||
[OpenAiChatEnum.GPT35]: {
|
||||
countTokens: ({ messages }: CountTokenType) =>
|
||||
countOpenAIToken({ model: OpenAiChatEnum.GPT35, messages }),
|
||||
adaptChatMessages: adaptChatItem_openAI
|
||||
},
|
||||
[OpenAiChatEnum.GPT4]: {
|
||||
countTokens: ({ messages }: CountTokenType) =>
|
||||
countOpenAIToken({ model: OpenAiChatEnum.GPT4, messages }),
|
||||
adaptChatMessages: adaptChatItem_openAI
|
||||
},
|
||||
[OpenAiChatEnum.GPT432k]: {
|
||||
countTokens: ({ messages }: CountTokenType) =>
|
||||
countOpenAIToken({ model: OpenAiChatEnum.GPT432k, messages }),
|
||||
adaptChatMessages: adaptChatItem_openAI
|
||||
}
|
||||
};
|
||||
|
||||
export const sliceTextByToken = ({
|
||||
model = 'gpt-3.5-turbo',
|
||||
text,
|
||||
length
|
||||
}: {
|
||||
model: ChatModelType;
|
||||
text: string;
|
||||
length: number;
|
||||
}) => {
|
||||
const enc = getOpenAiEncMap()[model];
|
||||
const encodeText = enc.encode(text);
|
||||
const decoder = new TextDecoder();
|
||||
return decoder.decode(enc.decode(encodeText.slice(0, length)));
|
||||
};
|
||||
106
src/utils/chat/openai.ts
Normal file
106
src/utils/chat/openai.ts
Normal file
@@ -0,0 +1,106 @@
|
||||
import { encoding_for_model, type Tiktoken } from '@dqbd/tiktoken';
|
||||
import type { ChatItemSimpleType } from '@/types/chat';
|
||||
import { ChatRoleEnum } from '@/constants/chat';
|
||||
import { ChatCompletionRequestMessage, ChatCompletionRequestMessageRoleEnum } from 'openai';
|
||||
|
||||
import Graphemer from 'graphemer';
|
||||
|
||||
const textDecoder = new TextDecoder();
|
||||
const graphemer = new Graphemer();
|
||||
|
||||
export const adaptChatItem_openAI = ({
|
||||
messages
|
||||
}: {
|
||||
messages: ChatItemSimpleType[];
|
||||
}): ChatCompletionRequestMessage[] => {
|
||||
const map = {
|
||||
[ChatRoleEnum.AI]: ChatCompletionRequestMessageRoleEnum.Assistant,
|
||||
[ChatRoleEnum.Human]: ChatCompletionRequestMessageRoleEnum.User,
|
||||
[ChatRoleEnum.System]: ChatCompletionRequestMessageRoleEnum.System
|
||||
};
|
||||
return messages.map((item) => ({
|
||||
role: map[item.obj] || ChatCompletionRequestMessageRoleEnum.System,
|
||||
content: item.value || ''
|
||||
}));
|
||||
};
|
||||
|
||||
/* count openai chat token*/
|
||||
let OpenAiEncMap: Record<string, Tiktoken>;
|
||||
export const getOpenAiEncMap = () => {
|
||||
if (OpenAiEncMap) return OpenAiEncMap;
|
||||
OpenAiEncMap = {
|
||||
'gpt-3.5-turbo': encoding_for_model('gpt-3.5-turbo', {
|
||||
'<|im_start|>': 100264,
|
||||
'<|im_end|>': 100265,
|
||||
'<|im_sep|>': 100266
|
||||
}),
|
||||
'gpt-4': encoding_for_model('gpt-4', {
|
||||
'<|im_start|>': 100264,
|
||||
'<|im_end|>': 100265,
|
||||
'<|im_sep|>': 100266
|
||||
}),
|
||||
'gpt-4-32k': encoding_for_model('gpt-4-32k', {
|
||||
'<|im_start|>': 100264,
|
||||
'<|im_end|>': 100265,
|
||||
'<|im_sep|>': 100266
|
||||
})
|
||||
};
|
||||
return OpenAiEncMap;
|
||||
};
|
||||
export function countOpenAIToken({
|
||||
messages,
|
||||
model
|
||||
}: {
|
||||
messages: ChatItemSimpleType[];
|
||||
model: 'gpt-3.5-turbo' | 'gpt-4' | 'gpt-4-32k';
|
||||
}) {
|
||||
function getChatGPTEncodingText(
|
||||
messages: { role: 'system' | 'user' | 'assistant'; content: string; name?: string }[],
|
||||
model: 'gpt-3.5-turbo' | 'gpt-4' | 'gpt-4-32k'
|
||||
) {
|
||||
const isGpt3 = model === 'gpt-3.5-turbo';
|
||||
|
||||
const msgSep = isGpt3 ? '\n' : '';
|
||||
const roleSep = isGpt3 ? '\n' : '<|im_sep|>';
|
||||
|
||||
return [
|
||||
messages
|
||||
.map(({ name = '', role, content }) => {
|
||||
return `<|im_start|>${name || role}${roleSep}${content}<|im_end|>`;
|
||||
})
|
||||
.join(msgSep),
|
||||
`<|im_start|>assistant${roleSep}`
|
||||
].join(msgSep);
|
||||
}
|
||||
function text2TokensLen(encoder: Tiktoken, inputText: string) {
|
||||
const encoding = encoder.encode(inputText, 'all');
|
||||
const segments: { text: string; tokens: { id: number; idx: number }[] }[] = [];
|
||||
|
||||
let byteAcc: number[] = [];
|
||||
let tokenAcc: { id: number; idx: number }[] = [];
|
||||
let inputGraphemes = graphemer.splitGraphemes(inputText);
|
||||
|
||||
for (let idx = 0; idx < encoding.length; idx++) {
|
||||
const token = encoding[idx]!;
|
||||
byteAcc.push(...encoder.decode_single_token_bytes(token));
|
||||
tokenAcc.push({ id: token, idx });
|
||||
|
||||
const segmentText = textDecoder.decode(new Uint8Array(byteAcc));
|
||||
const graphemes = graphemer.splitGraphemes(segmentText);
|
||||
|
||||
if (graphemes.every((item, idx) => inputGraphemes[idx] === item)) {
|
||||
segments.push({ text: segmentText, tokens: tokenAcc });
|
||||
|
||||
byteAcc = [];
|
||||
tokenAcc = [];
|
||||
inputGraphemes = inputGraphemes.slice(graphemes.length);
|
||||
}
|
||||
}
|
||||
|
||||
return segments.reduce((memo, i) => memo + i.tokens.length, 0) ?? 0;
|
||||
}
|
||||
|
||||
const adaptMessages = adaptChatItem_openAI({ messages });
|
||||
|
||||
return text2TokensLen(getOpenAiEncMap()[model], getChatGPTEncodingText(adaptMessages, model));
|
||||
}
|
||||
@@ -1,6 +1,6 @@
|
||||
import mammoth from 'mammoth';
|
||||
import Papa from 'papaparse';
|
||||
import { getEncMap } from './tools';
|
||||
import { getOpenAiEncMap } from './chat/openai';
|
||||
|
||||
/**
|
||||
* 读取 txt 文件内容
|
||||
@@ -154,7 +154,7 @@ export const splitText_token = ({
|
||||
maxLen: number;
|
||||
slideLen: number;
|
||||
}) => {
|
||||
const enc = getEncMap()['gpt-3.5-turbo'];
|
||||
const enc = getOpenAiEncMap()['gpt-3.5-turbo'];
|
||||
// filter empty text. encode sentence
|
||||
const encodeText = enc.encode(text);
|
||||
|
||||
|
||||
@@ -1,33 +1,5 @@
|
||||
import crypto from 'crypto';
|
||||
import { useToast } from '@/hooks/useToast';
|
||||
import { encoding_for_model, type Tiktoken } from '@dqbd/tiktoken';
|
||||
import Graphemer from 'graphemer';
|
||||
import type { ChatModelType } from '@/constants/model';
|
||||
|
||||
const textDecoder = new TextDecoder();
|
||||
const graphemer = new Graphemer();
|
||||
let encMap: Record<string, Tiktoken>;
|
||||
export const getEncMap = () => {
|
||||
if (encMap) return encMap;
|
||||
encMap = {
|
||||
'gpt-3.5-turbo': encoding_for_model('gpt-3.5-turbo', {
|
||||
'<|im_start|>': 100264,
|
||||
'<|im_end|>': 100265,
|
||||
'<|im_sep|>': 100266
|
||||
}),
|
||||
'gpt-4': encoding_for_model('gpt-4', {
|
||||
'<|im_start|>': 100264,
|
||||
'<|im_end|>': 100265,
|
||||
'<|im_sep|>': 100266
|
||||
}),
|
||||
'gpt-4-32k': encoding_for_model('gpt-4-32k', {
|
||||
'<|im_start|>': 100264,
|
||||
'<|im_end|>': 100265,
|
||||
'<|im_sep|>': 100266
|
||||
})
|
||||
};
|
||||
return encMap;
|
||||
};
|
||||
|
||||
/**
|
||||
* copy text data
|
||||
@@ -79,75 +51,3 @@ export const Obj2Query = (obj: Record<string, string | number>) => {
|
||||
}
|
||||
return queryParams.toString();
|
||||
};
|
||||
|
||||
/* 格式化 chat 聊天内容 */
|
||||
function getChatGPTEncodingText(
|
||||
messages: { role: 'system' | 'user' | 'assistant'; content: string; name?: string }[],
|
||||
model: 'gpt-3.5-turbo' | 'gpt-4' | 'gpt-4-32k'
|
||||
) {
|
||||
const isGpt3 = model === 'gpt-3.5-turbo';
|
||||
|
||||
const msgSep = isGpt3 ? '\n' : '';
|
||||
const roleSep = isGpt3 ? '\n' : '<|im_sep|>';
|
||||
|
||||
return [
|
||||
messages
|
||||
.map(({ name = '', role, content }) => {
|
||||
return `<|im_start|>${name || role}${roleSep}${content}<|im_end|>`;
|
||||
})
|
||||
.join(msgSep),
|
||||
`<|im_start|>assistant${roleSep}`
|
||||
].join(msgSep);
|
||||
}
|
||||
function text2TokensLen(encoder: Tiktoken, inputText: string) {
|
||||
const encoding = encoder.encode(inputText, 'all');
|
||||
const segments: { text: string; tokens: { id: number; idx: number }[] }[] = [];
|
||||
|
||||
let byteAcc: number[] = [];
|
||||
let tokenAcc: { id: number; idx: number }[] = [];
|
||||
let inputGraphemes = graphemer.splitGraphemes(inputText);
|
||||
|
||||
for (let idx = 0; idx < encoding.length; idx++) {
|
||||
const token = encoding[idx]!;
|
||||
byteAcc.push(...encoder.decode_single_token_bytes(token));
|
||||
tokenAcc.push({ id: token, idx });
|
||||
|
||||
const segmentText = textDecoder.decode(new Uint8Array(byteAcc));
|
||||
const graphemes = graphemer.splitGraphemes(segmentText);
|
||||
|
||||
if (graphemes.every((item, idx) => inputGraphemes[idx] === item)) {
|
||||
segments.push({ text: segmentText, tokens: tokenAcc });
|
||||
|
||||
byteAcc = [];
|
||||
tokenAcc = [];
|
||||
inputGraphemes = inputGraphemes.slice(graphemes.length);
|
||||
}
|
||||
}
|
||||
|
||||
return segments.reduce((memo, i) => memo + i.tokens.length, 0) ?? 0;
|
||||
}
|
||||
export const countChatTokens = ({
|
||||
model = 'gpt-3.5-turbo',
|
||||
messages
|
||||
}: {
|
||||
model?: ChatModelType;
|
||||
messages: { role: 'system' | 'user' | 'assistant'; content: string }[];
|
||||
}) => {
|
||||
const text = getChatGPTEncodingText(messages, model);
|
||||
return text2TokensLen(getEncMap()[model], text);
|
||||
};
|
||||
|
||||
export const sliceTextByToken = ({
|
||||
model = 'gpt-3.5-turbo',
|
||||
text,
|
||||
length
|
||||
}: {
|
||||
model?: ChatModelType;
|
||||
text: string;
|
||||
length: number;
|
||||
}) => {
|
||||
const enc = getEncMap()[model];
|
||||
const encodeText = enc.encode(text);
|
||||
const decoder = new TextDecoder();
|
||||
return decoder.decode(enc.decode(encodeText.slice(0, length)));
|
||||
};
|
||||
|
||||
Reference in New Issue
Block a user