gptchat 0.0.8

A local ChatGPT interface

gptchat      

NOTE: Still in early development. Documentation not available yet.

pip install gptchat

gptchat makes it easy to develop with OpenAI's API, without having to build your own user interface.

This python package lets you instantly host a ChatGPT look-alike frontend application on your machine.

As the developer, your only responsibility is to define what happens on the server when the user sends a message.

When the user hits send, the conversation (a list of OpenAI Message objects) will be passed to your handler function. You have full control over what to send back, and when. While your handler function is executing, you may emit updates with full messages, or partial messages (deltas) if streaming.

Function calls (and soon, tool calls) are fully supported. The frontend will render function/tool calls for you, in the same style as ChatGPT.

Minimal learning curve necessary

gptchat's API is merely a reflection of the OpenAI API. Any data you send in the messages argument to OpenAI can also be passed to the gptchat frontend, where it will be rendered.

Getting Started

A minimal configuration to get you up and running.

from gptchat import app
from openai import OpenAI

openai_client = OpenAI(api_key="YOUR_API_KEY")

@app.handle_send_message
def message_handler(messages):
    # get response from openai
    response = openai_client.chat.completions.create(
        messages=messages,
        model="gpt-4",
    )
    new_msg = response.choices[0].message

    # send update to frontend
    app.update_message(new_msg)
    

app.run_app(debug=True, port=5002)

That doesn't look quite like ChatGPT yet. The response is sent all at once, after a long delay. Let's try streaming instead.

We'll define a utility function to stream an OpenAI response

def openai_chat_stream(messages, model: str = "gpt-4", **kwargs):
    """
    Streams an OpenAI chat completion message, and yields deltas.
    """
    stream = openai_client.chat.completions.create(
        messages=messages,
        model=model,
        stream=True,
        **kwargs
    )
    for chunk in stream:
        if chunk.choices:
            yield chunk.choices[0]

On the frontend, each message must have a unique identifier. In the previous example, when we ran app.update_message(new_msg), a unique id was was assigned to new_msg for us.

But when sending partial updates, we need to make sure each update has the same id, so they can be combined into a single message.

Here's a new version of message_handler() for streaming.

@app.handle_send_message
def message_handler(messages):
    stream = openai_chat_stream(messages)

    id = app.new_id()
    for chunk in stream:
        app.update_message(chunk.delta, id=id)

Optional Configuration

Configurations can be set using the set_config() function:

app.set_config({
    "some_option": "some value",
    "some_option2": "some other value",
})

The following options are available:

messages

Useful during development, set the app to start with some messages already in the chat.

app.set_config({
    "messages": [
        { "role": "user", "content": "Hello, how are you?" },
        { "role": "assistant", "content": "I'm doing great, how can I help you today?" },
    ]
})

functions

Specify display headers for function/tool calls

By default, function calls are displayed with a header, "Function call to your_function()":

The functions option lets you change what gets displayed here, for any functions.

The text you provide is rendered as markdown.

app.set_config({
    "functions": {
        "get_stock_price": "Fetching **stock price**...",
    }
})

Utilities

concat_stream()

Combine all the partial updates received from an OpenAI stream into a single message Choice.

An essential tool when streaming responses from OpenAI while using function/tool calling. Keep a record of all the accumulated updates received during a stream, and use concat_stream() to combine them into a single message Choice.

Let's rewrite our message_handler() function from the earlier example, to use concat_stream() so it can handle function calls.

from gptchat.utils import concat_stream

@app.handle_send_message
def message_handler(messages):
    stream = openai_chat_stream(messages)

    id = app.new_id()
    streamed_chunks = []
    for chunk in stream:
        app.update_message(chunk.delta, id=id)
        streamed_chunks.append(chunk)

    full_message = concat_stream(streamed_chunks)
    # Do something with the full message, like handle function calls

Notes for Contributors

This project has two parts:

1. /client: Frontend typescript application. All the core logic of gptchat runs here, in the browser. This builds a static bundle to the python package. (/package/build)
2. /package: A python package that acts as the plumbing for your server. It takes care of serving the client bundle, and provides wrapper functions for implementing your websocket communications.

Getting started

Client source: /client/src/lib. (Note, the bulk of the logic and state management is in main.ts.)

Python package source: /package/gptchat

• Run the frontend on a dev server: From /client, run npm install. To start the server, use npm run dev.
  • In production, there is only one server, so the client code uses an empty string as the URL path for the websocket connection, which makes a request from the current URL. But in development, the frontend runs on a separate Vite server. To connect to a separate Flask server running locally, visit the socket.ts module, switch PROD=false, and ensure the URL matches the port of your local flask server.
• Build the frontend: From /client, run npm run build. This should save the bundle to /package/build.
  • The python server expects this bundle directory to be present when the flask app starts. So it's recommended to do this first, even if you're just starting development, and even if your client bundle doesn't work. The build only gets served if you visit your Flask server's URL in the browser, which isn't necessary for development anyway.
• Install gptchat in editable mode: Once you've ran npm run build at least once, install the python package by running pip install -e . from inside /package, using whatever virtual environment you'll be developing out of.
  • You only have to do this once. An editable pip install means that import gptchat will always pull from the current state of the source code on your machine. So you can write your test app scripts anywhere on your machine, while directly making code changes in /gptchat