yoquGPT / w2uGPT / iwuGPT: Hacking ChatGPT, or just an use case to play for a while

What does yoqu, w2u, iwu mean?:

• yoqu : yo quiero usar
• w2u : want to use
• iwu : I want to use

A wish, an attempt to use some technologies, a technical challenge, an excuse to learn about something, whatever you want, ¯_(ツ)/¯_

What do you have in this repository?

First the idea:

I want to use a resource like ChatGPT or something similar to satisfy an use case (e.g.: an application using some LLM model), so I need:

• An API (e.g. FastAPI)
• An application (e.g. streamlit app, Langchain)
• A way to use this resource (e.g. RPA)
• Probably a DB

Video demo: demo.mp4

Disclaimer

• Note that this is not the way to use ChatGPT, or Claude or any resource
• A real application should use the API (e.g. https://platform.openai.com/docs/introduction)
• ChatGPT responses are not suitables for production use cases

After the first version of this development, OpenAI made several changes to the ChatGPT interface (and will probably continue doing so).

I like to say that some RPAs are similar to "sneaking into a party through the back door": you know that you should get an invitation or buy a ticket...

:-)

This POC/use case involve the following mainly:

• RPA with Selenium
• FastAPI
• Custom LLM (Langchain)
• Streamlit
• Some bash scripts

Installation and configuration

1. Install jq (bash scripts call it)

2. Install Chrome Browser and undetected_chromedriver

3. Install Python v3.10

4. Create a virtual env (recommended) and activate it

   python -m venv yoqu
   source yoqu/bin/activate
   pip install --upgrade pip
   

5. Clone/download this repository

   git clone https://github.com/ktxo/yoqu-gpt.git
   cd yoqu-gpt
   

6. Install dependencies:

   pip install -r requirements.txt
   

7. Configure resource configuration file

   Modifiy the following:

   • name
     • desc
     • dump_folder
     • command / options_experimental
     • chromedriver
   • uc_options
   • Copy undetected_chromedriver binary to the folder defined in chromedriver
   • Update port or path options if you want: port number (e.g. 9222) and folder user-data-dir) must be unique.

   Each RPA needs a Chrome profile with an already signed-in session and their credentials saved in the profile.

   This RPA doesn't solve any captcha (CloudFare), when RPA open the browser go to the url and expect to enter (don't do any login, I hate solving captchas and these kinds of things; they are boring, and I avoid them").

   Run:

   ./yoqu_configure.sh  <config_file.json>
   

   This script will execute Chrome and open the url, you need to login in, save user, password and then close the browser

   The second time you run the script you should see the chat page instead of login page

8. Configure RPA Manager configuration file rpa_manager.json with the list of resource files from previous step

9. Run poc:

   See images/demo.mp4

   • Start API executing run_api.sh

     • You should see the browser login to the resource (ChatGPT)
     • Swagger Yoqu API (http://0.0.0.0:8000/docs)

   Run one of these:

   • Streamlit app based on Build a ChatGPT-like app using Yoqu API

     run_app_st_chatgpt-like-clone-yoqu.sh

   • YoquChatGPTLLM Streamlit app with API/Langchain custom LLM using Yoqu API (YoquChatGPTLLM)

     run_app_st_langchain_llm.sh

Resource configuration file

Each YoquResource (a RPA) has a configuration file (see yoqu_rpa_chatgpt1.json or yoqu_rpa_claude.json) like:

{
  "desc": "Some description for this resource",
  "type": "chatgpt",      ## Type: chatgpt | claude
  "dump": true,           ## Enable to save data to dump folder 
  "dump_folder": "dump",  ## Dumpt folder
  
                          ## These options are used by each specific resource
  "resource": {
    "debug": false,                     ## Pending to allow dump page/screenshot      
    "url": "https://chat.openai.com",   ## URL for RPA
    "url_login": "https://openai.com",  ## URL to login to the site (used by yoqu_configure.sh)
    "url_title": "ChatGPT",             ## URL title for url resource.url
    
    ## Command to launch te browser, browser is started using : psutil.Popen([resource.command] + [resource.url])  
    "command": [                                
      "/opt/google/chrome/chrome",              ## Path to chrome browser
      "--remote-debugging-port=9222",           ## Debug port for this browser instance, must be unique 
      "--user-data-dir=chrome_data/chatgpt"     ## Path to chrome data 
    ],
    "sleep_range": [1,3],                       ## RPA random delays: time.sleep(random.randint(*sleep_range))
    "options": ["--log-level=3"],               ## Chromedriver options
    "options_experimental": [                   ## Chromedriver experimental options
      ["debuggerAddress", "127.0.0.1:9222"]
    ],
    "uc_options": {                                         ## undetected_chromedriver options
      "headless": false,
      "use_subprocess": true,
      "user_data_dir": "chrome_data/chatgpt2",              ## Path to chrome data
      "driver_executable_path": "./undetected_chromedriver" ## Path to undetected_chromedriver
    }    
  },
  "wait": 10
}

Next steps, improvements, changes, y otras yerbas

• Review all (too dirty)
• Some improvements:
  • Review error handling and retries in RPA, for example: ChatGPT send errors when there are high traffic in their platform
  • Detect and bypass Cloudflare captcha
  • Headless browser
  • Automatic configuration of Chrome Profile from configuration file
  • Add "ephemeral" chats: create chat, get response, delete chat (✅)
  • Review response time (since browser receive the response)
• Pending:
  • Add delete chat operation to API
  • Fix scroll all chats (RPAChatGPTResource_select_chat())
  • Review Custom LLM: output parsers
  • Process handling
  • Review
    • restart operation
    • Review stats/status
    • dump
• Add Claude (https://claude.ai/) (I'm testing it)
• Add more resources (e.g. Bard, Perplexity)
• DB design/CRUD
• Remote RPA instances (different machines)
• Add more use cases: spacy integration, more langchain
• ...

After the latest changes in the ChatGPT interface, I noticed weird errors controlling the page with Selenium. I deleted the Chrome profile to solve it.

Final

If you find this useful, interesting, boring, unuseful, funny, waste of time, or whatever feel free to contact me