doschgpt

A proof-of-concept ChatGPT and Hugging Face client for DOS.

Photos of the client running on my 1984 IBM 5155 Portable PC with a 4.77Mhz Intel 8088 CPU with MS-DOS 6.22.

As there are no native HTTPS APIs for DOS, a HTTP-to-HTTPS proxy like this I've written running on a modern machine is needed.

This program is heavily based on sample code in the DOS networking MTCP library. The program also requires a DOS Packet Driver to be loaded and MTCP to be set for the machine/VM.

Using the Text-to-speech feature requires a Sound Blaster-compatible card and 8Mhz CPU.

To reduce coding complexity and memory usage, the app only sends the previous request/reply set together with the current request to provide the conversational context to the model instead of the entire conversation history. Do notify me if you believe this is an issue for you and you prefer more history to be sent.

This program was written in a short time as a toy project. It has not been vigorously tested thus is NOT meant for "production" use.

Using the application

Application binary can be found in the releases directory or Github Releases section but do the following first.

Both OpenAI and Hugging Face requires an API key/token to use their APIs. Follow the instructions on OpenAI or Hugging Face websites to obtain this key before proceeding. A read token for Hugging Face is sufficient.
Download and start up http-to-https-proxy on a modern machine/SBC.
The application requires a config file. By default it will use the filename doschgpt.ini but you can specify another path with the -c argument. Modify the configuration file to suit your needs in this order. A sample file can be found with the binary.

API key: Place your key without quotes (API key in this sample file has been revoked)
Model: Language model to use such as but not limited to. gpt-3.5-turbo for ChatGPT. facebook/blenderbot-400M-distill or microsoft/DialoGPT-large for Hugging Face. More details and models for Hugging Face can be found here.
Request Temperature: How random the completion will be. More OpenAI details and Hugging Face details
Proxy hostname: Hostname IP of the proxy
Proxy port: Proxy Port
Outgoing start port: Start of a range of randomly selected outgoing port
Outgoing end port: End of a range of randomly selected outgoing port
Socket connect timeout (ms): How long to wait when attempting to connect to proxy
Socket response timeout (ms): How long to wait for OpenAI's servers to reply

Ensure that your DOS environment has loaded the following:

Packet Driver
MTCP Config Environment variable MTCPCFG
MTCP Config file configured by DHCP
Text-to-speech feature requires the BLASTER variable such as SET BLASTER=A220 I5 T1 D1 T4 to be set.

Just launch doschgpt.exe in your machine and fire away. Press the ESC key to quit the application. You may use the following optional command line arguments.

-hf: To use Hugging Face instead of ChatGPT
-cdoschgpt.ini: Replace doschgpt.ini with any other config filepath you desire. There is no space between the -c and the filepath.
-dri: Print the outgoing port, number of prompt and completion tokens used after each request. Tokens are only provided by ChatGPT.
-drr: Display the raw server return headers and json reply
-drt: Display the timestamp of the latest request/reply
-cp737: Supports Greek Code Page 737. Ensure code page is loaded before starting the program.
-fhistory.txt: Append conversation history to new/existing text file. File will also include debug messages if specified. Replace history.txt with any other filepath you desire. There is no space between the -f and the filepath.
-sbtts: Able to read server reply using a text-to-speech driver used by Dr. Sbaitso.

Example usage:

# Connects to ChatGPT using config file at doschgpt.ini
doschgpt.exe

# Connects to Hugging Face using config file at hf.ini with timestamps after each request and reply
doschgpt.exe -hf -chf.ini -drt

Parsed options will be displayed.

Release details

doschgpt.exe: Main binary
doschgpt.ini: Sample configuration file for ChatGPT
hf.ini: Sample configuration file for Hugging Face
Sbtalker.exe: Smoothtalker by First Byte text-to-speech engine that loads as a TSR (This is called by the client on start)
Blaster.drv: Used by Smoothtalker to talk to a Sound Blaster card
Remove.exe: Unloads the Smoothtalker TSR (This is called by the client on end)
Read.exe: Reads its command line arguments to the Smoothtalker TSR

Compilation

To compile this application, you have to use Open Watcom 2.0 beta which you can download from here. Open Watcom 2.0 for 64-bit Windows which was released on 2023-04-01 02:52:44 is used. The v1.9 version seems to create binaries with issues on some platforms.

During installation, Open Watcom may prompt to install the environment variables. I chose to not do that to avoid having those variables being permanent. Instead I use a batch file to set the variables whenever I need to compile.

The program is compiled via a Makefile that is referenced from MTCP.

# Open cmd.exe
cd doschgpt-code

# If using Open Watcom v2.0 beta installed to C:\WATCOM2
20setenv.bat

# If using Open Watcom v1.9 installed to C:\WATCOM (Not recommended)
19setenv.bat

# To compile
wmake

#  Only if using Open Watcom 1.9. To patch the Open Watcom runtime to support Compaq Portable. Not needed for Open Watcom 2.0 beta.
PTACH.exe doschgpt.exe doschgpt.map -ml

# To clean
wmake clean

This application compiles against the MTCP library. I have unzipped the latest version mTCP-src_2023-03-31.zip at the time of development to the mtcpsrc directory. When Brutman updates this library again in future, simply replace the contents of the mtcpsrc directory with the new library.

PTACH.exe is a Win NT program compiled from MTCP sources.

Development

I use Visual Studio Code text editor to code for ease of use. For ease of testing, I used a virtual machine to run the application as 16-bit DOS applications and the MTCP network stack cannot run on modern Windows.

More details of my setup can be found here.

To easily transfer the binary, I used Python to host my build directory as a temporary webserver. Then use the MTCP tool htget to fetch the binary.

# On modern machine with binary
python3 -m http.server 8000

# Run on DOS machine/VM
htget -o doschgpt.exe http://X.X.X.X:8000/doschgpt.exe

Mock proxy

OpenAI implements rate limits on their API hence we should minimise calling their API repeatedly.

To avoid calling the OpenAI's servers during testing, we can mock the server's using this mockprox.goGo program that will replay the contents of reply.txt whenever the API call is received.

cp correct.txt reply.txt
go build mockprox.go
mockprox.exe

APIs

# Test API directly
curl https://api.openai.com/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer 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" -d '{ "model": "gpt-3.5-turbo", "messages": [{"role": "user", "content": "What is MS-DOS?"}], "temperature": 0.7 }'

# Call through the https proxy for testing
curl --proxy "http://192.168.1.144:8080" https://api.openai.com/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer sk-EhmTsEsKyH4qHZL2mr3hT3BlbkFJd6AcfdBrujJsBBGPRcQ" -d '{ "model": "gpt-3.5-turbo", "messages": [{"role": "user", "content": "What is MS-DOS?"}], "temperature": 0.7 }'

# Hugging Face API
curl https://api-inference.huggingface.co/models/facebook/blenderbot-400M-distill -X POST -H "Authorization: Bearer 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" -d '{"inputs": {"text":"What is retro-computing?"}, "parameters": { "temperature": 1.0 } }'

# Hugging Face API with history
curl https://api-inference.huggingface.co/models/facebook/blenderbot-400M-distill -X POST -H "Authorization: Bearer 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" -d '{"inputs": {"past_user_inputs": ["What is retrocomputing?"], "generated_responses": ["A retro computer is a type of computer that was invented in the 1950s."], "text": "I want one", "parameters": { "temperature": 1.0 } }'

Changelog

Go to CHANGELOG.md.

yeokm1/doschgpt

yeokm1

Reviews

Repository Details