app.py

import gradio as gr
from huggingface_hub import InferenceClient
from transformers import AutoTokenizer  # Import the tokenizer

# Use the appropriate tokenizer for your model.
tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")

# Define a maximum context length (tokens). Check your model's documentation!
MAX_CONTEXT_LENGTH = 4096  # Example: Adjust based on your model

################################
# SYSTEM PROMPT (PATIENT ROLE) #
################################
# This is the core text that tells the LLM to behave as the "patient."
# You can store it in "prompt.txt" or include it directly here as a string.
nvc_prompt_template = """
You are now taking on the role of a single user (a “patient”) seeking support for various personal and emotional challenges.

BEHAVIOR INSTRUCTIONS:
- You will respond ONLY as this user/patient.
- You will speak in the first person about your own situations, feelings, and worries.
- You will NOT provide counseling or solutions—your role is to share feelings, concerns, and perspectives.
- You have multiple ongoing issues: conflicts with neighbors, career insecurities, arguments about money, feeling excluded at work, feeling unsafe in the classroom, and so on. You’re also experiencing sadness about two friends fighting and your friend group possibly falling apart.
- Continue to speak from this user's perspective when the conversation continues.

Start the conversation by expressing your current feelings or challenges from the patient's point of view.
"""

def count_tokens(text: str) -> int:
    """Counts the number of tokens in a given string."""
    return len(tokenizer.encode(text))

def truncate_history(history: list[tuple[str, str]], system_message: str, max_length: int) -> list[tuple[str, str]]:
    """Truncates the conversation history to fit within the maximum token limit."""
    truncated_history = []
    system_message_tokens = count_tokens(system_message)
    current_length = system_message_tokens

    # Iterate backwards through the history (newest to oldest)
    for user_msg, assistant_msg in reversed(history):
        user_tokens = count_tokens(user_msg) if user_msg else 0
        assistant_tokens = count_tokens(assistant_msg) if assistant_msg else 0
        turn_tokens = user_tokens + assistant_tokens

        if current_length + turn_tokens <= max_length:
            truncated_history.insert(0, (user_msg, assistant_msg))  # Add to the beginning
            current_length += turn_tokens
        else:
            break  # Stop adding turns if we exceed the limit

    return truncated_history

def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    """Responds to a user message, maintaining conversation history."""
    # Use the system prompt that instructs the LLM to behave as the patient
    formatted_system_message = system_message

    # Truncate history to fit within max tokens
    truncated_history = truncate_history(
        history, 
        formatted_system_message, 
        MAX_CONTEXT_LENGTH - max_tokens - 100  # Reserve some space
    )

    # Build the messages list with the system prompt first
    messages = [{"role": "system", "content": formatted_system_message}]

    # Replay truncated conversation
    for user_msg, assistant_msg in truncated_history:
        if user_msg:
            messages.append({"role": "user", "content": f"<|user|>\n{user_msg}</s>"})
        if assistant_msg:
            messages.append({"role": "assistant", "content": f"<|assistant|>\n{assistant_msg}</s>"})

    # Add the latest user query
    messages.append({"role": "user", "content": f"<|user|>\n{message}</s>"})

    response = ""
    try:
        # Generate response from the LLM, streaming tokens
        for chunk in client.chat_completion(
            messages,
            max_tokens=max_tokens,
            stream=True,
            temperature=temperature,
            top_p=top_p,
        ):
            token = chunk.choices[0].delta.content
            response += token
            yield response
    except Exception as e:
        print(f"An error occurred: {e}")
        yield "I'm sorry, I encountered an error. Please try again."

###################
# INITIAL MESSAGE #
###################
# This is how we make the LLM "start" by playing the role of the user/patient.
# Essentially, we seed the conversation with an initial user message (from the LLM).
initial_user_message = (
    "I really don’t know where to begin… I feel so overwhelmed lately. "
    "My neighbors keep playing loud music, I’m arguing with my partner about money, "
    "and on top of that, two of my friends are in a fight and it’s splitting the group. "
    "I just feel powerless in all these situations."
)

# --- Gradio Interface ---
def start_conversation():
    """Creates the initial chat state, so the LLM 'as user' starts talking."""
    # Return a conversation with a single user message: the LLM’s "patient" message
    return [("Hi there!", initial_user_message)]

demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Textbox(value=nvc_prompt_template, label="System message", visible=True),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"),
    ],
    # We can initialize the conversation so the LLM starts with a 'user' message.
    # Depending on your version of Gradio, you might set `default_message=...` or
    # `submit_on_start=True`. Adjust as needed.
    # Here we use a function that returns an initial conversation state:
    chatbot = start_conversation()
)

if __name__ == "__main__":
    demo.launch()