jupyter-anywidget-window-ai 0.0.2

No project description provided

jupyter_anywidget_window_ai

Example of using window.ai in-browser LLM (Chrome Canary) in JupyterLab / Jupyter notebook using anywidget.

Set-up Chrome Canary / window.ai: Getting started with window.ai in Chrome Canary — in browser Gemini LLM

We can test the availability of window.ai prompt API using some simple javascript magic in a Jupyter notebook accessed via an appropriately configured Chrome Canary browser:

%%javascript

// h/t ChatGPT

(async function() {
     console.log("Creating AI asisstant...\n\n");
 
    // Assuming ai.assistant.create() is defined 
    // and accessible in your environment
    const session = await ai.assistant.create();

    console.log("Session created...\n\n");

    console.log("Sending prompt...\n\n");

    // Prompt the model and wait for the result
    const result = await session.prompt("Write me a poem.");


    // Log the result to the console
    console.log(result);
})();

Installation

pip install jupyter_anywidget_window_ai

Usage

This runs in a browser based Jupyter environment and uses the browser machinery to run the in-browser LLM application.

Create a widget:

# Import package
from jupyter_anywidget_window_ai import windowai_panel, windowai_inline, windowai_headless

# Create a headless widget
# - works in: Jupyter Lab, Jupyter Notebook, VS Code
w = windowai_headless()

# Inline mode
# Preview output as cell output from initialising cell
# - works in: Jupyter Lab, Jupyter Notebook, VS Code
# w = windowai_inline()

# Create a widget panel with a handle
# - uses jupyter sidecar (Jupyter Lab only)
#w = windowai_panel()
# Close panel either via X or:
#w.close()

Create a prompt session with an optional system prompt: w.create_session(system_prompt=PROMPT)

Submit a prompt: w.prompt("Hello, how are you?")

Also available: w.create_session(..., initialPrompts=[] )

At the moment, there is no callback available on a response ggenerated from a prompt. However, an audible alert is available.

This is enabled by default in windowai_headless(). The response can also be spoken. Use boolean (True/False ) parameters completion_tone= and speak_msg= in windowai_panel(), windowai_inline() and windowai_headless() to set audible responses.

Get the response text: w.get_latest_output()

Destroy the session: w.destroy()

See window.ai.assistant docs for more examples that should translate: Google explainer: prompt-api

TO DO

• add support for changing temperature, topK;
• add autosession flag to allow session to be automatically configured and created when a widget is instantiated;
• add support for other APIs (explainers-by-googlers/writing-assistance-apis)