Skip to content
This repository has been archived by the owner on Sep 3, 2023. It is now read-only.

Latest commit

 

History

History
141 lines (99 loc) · 7.98 KB

README.en.md

File metadata and controls

141 lines (99 loc) · 7.98 KB
Raw

Ex-ChatGPT - ChatGPT with ToolFormer

language GitHub GitHub last commit GitHub Repo stars

简体中文 English / Background

ChatGPT can act as a tool former without requiring adjustment, generating API requests for questions to assist in answering. Ex-ChatGPT enables ChatGPT to call external APIs, such as WolframAlpha, Google, and WikiMedia, to provide more accurate and timely answers.

This project is divided into Ex-ChatGPT and WebChatGPTEnhance. The former is a service that uses the GPT3.5 Turbo API and Google,WolframAlpha,WikiMedia APIs, while the latter is a browser extension which update the origin WebChatGPT plugin to Enable adding external API, supportting ChatGPT webpage to call different APIs and prompts

user interface display

ExChatGPT

chatHistory

WebChatGPTEnhance

WebChatGPT

Highlights

  • OAuth2.0 Multi User Management (In webTest branch)
  • Speech Dialogue Functionality, utilizing Microsoft Azure API to optimize response time (around 1-2 seconds), featuring both speech recognition and text-to-speech capabilities, supporting multiple languages and voices, and allowing for customized voices.
  • Docker and Proxy support
  • Redundant backup of chat history
  • Support for OpenAI GPT-3.5 Turbo API
  • Ability for ChatGPT to call external API interfaces, such as Google, WolframAlpha, and WikiMedia
  • Clean up Google search results data to reduce token usage.
  • Automatic saving and loading of conversation history with automatic compression
  • Ability to display the number of tokens used
  • API pool
  • Markdown and MathJax rendering
  • API call process displayed with an animation, similar to Bing
  • Conversation management with load and chat modes similar to ChatGPT web page layout
  • Shortcut keys for quickly selecting modes (Tab), line breaks (Shift+Enter), sending messages (Enter), and selecting message history (up and down), similar to a terminal.
  • stream feature is similar to that of a typewriter, it responds faster with results instead of loading everything at once. Stream outputs the results in steps, as shown in the example: stream
  • Automatic prompt completion in Chat Mode with support for fuzzy search, pinyin search, and custom prompt selection.The project comes with prompt from awesome-chatgpt-prompts. promptCompletion

To Do List

  • Mobile interface adaptation
  • OCR recognition of formula text in images sent by users
  • OAuth2.0 multi-user authentication
  • Call the diffusing model to generate images (achieving similar multimodal effects)
  • Further spider summary and data cleansing of web search results
  • Add code running API, and more APIs
  • Embedding alignment retrieval of chat records/local knowledge databases

Installation

Ex-chatGPT

  • pip install pip install -r requirements.txt
  • Copy apikey.ini.example and rename it as apikey.ini. Then, fill in your API key and proxy in apikey.ini. If you only have one OpenAI API key, you should delete key1 = sk-xxxx；key2 = sk-xxxx.
    • Google api key and search engine id apply
    • wolframAlpha app id key apply
    • openAI api key(new feature) or chatGPT access_token(old version) apply
    • (optional, Text To Speech And Speech recognition) fill in Azure API key and region apply
  • run the main.py and click the local url like http://127.0.0.1:1234/
  • change the mode in the selection box, now have chat,detail,web,webDirect,WebKeyWord
  • Voice Conversation Chat(optional feature), select language and voice in chatGPTEx/static/styles/tts.js, click the microphone on chat interface to start/close conversation mode.

WebChatGPTEnhanceExtension

  • fill you Googgle api key and client id in chatGPTChromeEhance/src/util/apiManager.ts/getDefaultAPI
  • run npm install
  • run npm run build-prod
  • get the extension in chatGPTChromeEhance/build
  • add your prompts and APIs in option page.
    • APIs and prompts examples are in /WebChatGPTAPI
    • wolframAlpha needs to run local sever - WebChatGPTAPI/WolframLocalServer.py

Mode Introduction

Web Mode

The Web Mode starts by asking ChatGPT a question directly. ChatGPT generates a series of API calls related to the query and uses the first returned result and question to verify and supplement with an API call. Finally, ChatGPT summarizes the information. Web Mode has a better chat ability than just summarizing the response.

Chat Mode

Chat Mode only calls the OpenAI API interface, similar to the web version of ChatGPT. You can search and choose different prompts by typing /promtname, and it supports fuzz search.

WebDirect Mode

WebDirect Mode first lets ChatGPT generate a series of API calls related to the query. Then, it directly calls a third-party API to search for the answer to each query, and finally, ChatGPT summarizes the information. WebDirect Mode is faster and relatively more accurate for single query information.

Detail Mode

Detail Mode is an extension of the WebDirect Mode, where an additional API call is made to supplement the current results with further inquiries (such as information not found in the previous search). Finally, ChatGPT summarizes the information.

Keyword Mode

Keyword Mode generates keywords directly from ChatGPT for querying, using DDG. It doesn't require other API keys. However, its accuracy is relatively poor.

Update Log

  • Clean up Google search results data to reduce token usage.

  • Update all API proxy pools and add API restriction cooldown mechanism (Google 403 cooldown for 1 day).

  • Voice dialogue function, using Microsoft Azure API, optimized response speed, including voice recognition and text-to-speech, supporting multiple voices and languages, custom voice.

  • stream feature is similar to that of a typewriter, it responds faster with results instead of loading everything at once. Stream outputs the results in steps, as shown in the example: stream

  • Redundant backup of chat history.

  • Auto-complete prompt selection in chat mode with fuzzy search and Pinyin search support. promptCompletion

  • Update docker and proxy support.

  • Shortcut keys for quick mode selection (Tab) and line break (Shift + Enter), while Enter sends the message. Message history selection (up, down) is similar to a terminal.

  • Update chat history management sidebar. chatHistory

  • Update API calls processing animation. APIAnimation

  • Web page beautification. WebBeautification

  • Update Markdown and MathJax renderer. MathJax

  • Update chat history token optimizer, and web mode can respond according to chat history. Add token cost counter. history

  • Update web chat mode selection in the webpage and optimize prompt and token cost. Restrict the token limit. mode

  • Update better support for Chinese queries and add current date info. date

  • Update web chat mode and fix some bugs.

  • Update API configuration.

  • Update API pool.

  • Automatic saving and loading of conversation history, ChatGPT can retrieve previous conversations.

  • Update official support for OpenAI GPT3.5 Turbo API, which is super fast and cheap.

  • Update extra API calls and search summarizations to provide a more comprehensive and detailed answer.