app.py

import os
os.system("pip uninstall -y gradio")
os.system("pip install gradio==3.50.2")

from huggingface_hub import InferenceClient
import gradio as gr

"""
Chat engine.

TODOs:
- Better prompts.
- Output reader / parser.
- Agents for evaluation and task planning / splitting.
    * Haystack for orchestration
- Tools for agents
    * Haystack for orchestration
- 

"""

selected_model = "mistralai/Mixtral-8x7B-Instruct-v0.1"

client = InferenceClient(selected_model)

def format_prompt(query, history, lookback):
    prompt = "Responses should be no more than 100 words long.\n"
    
    for previous_query, prevous_completion in history[-lookback:]:
        prompt += f"<s>[INST] {previous_query} [/INST] {prevous_completion}</s> "
    
    prompt += f"[INST] {query} [/INST]"
  
    return prompt
    
def query_submit(user_message, history):
    return "", history + [[user_message, None]]

def query_completion(
    query,
    history,
    lookback = 3,
    max_new_tokens = 256,
):

    generateKwargs = dict(
        max_new_tokens = max_new_tokens,
        seed = 1337,
    )

    formatted_query = format_prompt(query, history, lookback)
    
    stream = client.text_generation(
        formatted_query,
        **generateKwargs,
        stream = True,
        details = True,
        return_full_text = False
    )
    
    history[-1][1] = ""
    
    for response in stream:
        history[-1][1] += response.token.text
        yield history

def retry_query(
    history,
    lookback = 3,
    max_new_tokens = 256,
):
    if not history:
        pass
    
    else:
        query = history[-1][0]
        history[-1][1]  = None
        
        generateKwargs = dict(
            max_new_tokens = max_new_tokens,
            seed = 1337,
        )
        
        formatted_query = format_prompt(query, history, lookback)
        
        stream = client.text_generation(
            formatted_query,
            **generateKwargs,
            stream = True,
            details = True,
            return_full_text = False
        )
        
        history[-1][1] = ""
        
        for response in stream:
            history[-1][1] += response.token.text
            yield history

        
"""
Chat UI using Gradio Blocks.

Blocks preferred for "lower-level" layout control and state management.

TODOs:
- State management for dynamic components update.
- Add scratchpad readout to the right of chat log.
    * Placeholder added for now.
- Add functionality to retry button.
    * Placeholder added for now.
- Add dropdown for model selection.
    * Placeholder added for now.
    
"""

with gr.Blocks() as chatUI:
    # gr.State()
    
    with gr.Row():
        modelSelect = gr.Dropdown(
            label = "Model selection:",
            scale = 0.5,
        )
    
    with gr.Row():
        chatOutput = gr.Chatbot(
            bubble_full_width = False,
            scale = 2
        )
        agentWhiteBoard = gr.Markdown(scale = 1)
        
    with gr.Row():
        queryInput = gr.Textbox(
            placeholder = "Please enter you question or request here...",
            show_label = False,
            scale = 4,
        )
        submitButton = gr.Button("Submit", scale = 1)
        
    with gr.Row():
        fileUpload = gr.File(
            height = 100,
        )
        retryButton = gr.Button("Retry")
        clearButton = gr.ClearButton([queryInput, chatOutput])
    
    with gr.Row():
        with gr.Accordion(label = "Expand for edit system prompt:"):
            systemPrompt = gr.Textbox(
                value = "System prompt here (null)",
                show_label = False,
                lines = 4,
                scale = 4,
        )
    

    """
    Event functions
    
    """
    queryInput.submit(
        fn = query_submit,
        inputs = [queryInput, chatOutput],
        outputs = [queryInput, chatOutput],
        queue = False,
    ).then(
        fn = query_completion,
        inputs = [queryInput, chatOutput],
        outputs = [chatOutput],
    )
    
    submitButton.click(
        fn = query_submit,
        inputs = [queryInput, chatOutput],
        outputs = [queryInput, chatOutput],
        queue = False,
    ).then(
        fn = query_completion,
        inputs = [queryInput, chatOutput],
        outputs = [chatOutput],
    )

    retryButton.click(
        fn = retry_query,
        inputs = [chatOutput],
        outputs = [chatOutput],
    )

    
chatUI.queue()
chatUI.launch(show_api = False)