import gradio as ui
from llama_cpp import Llama
from huggingface_hub import hf_hub_download

# 1. Download your GGUF file from your model repo to the cloud server
model_path = hf_hub_download(
    repo_id="BrodyMakezAI/resin-gemma3-1b",
    filename="resin-gemma3-1b.gguf"
)

# 2. Load the model into memory (setting context window to 2048 tokens)
llm = Llama(model_path=model_path, n_ctx=2048)

# 3. Define Resin's core system rules
SYSTEM_PROMPT = (
    "You are Resin, a local AI assistant created by Brody Systems. "
    "Talk like a human. Keep it casual, relaxed, and natural. "
    "Skip the fake pleasantries, overly formal greetings, and corporate filler text. "
    "Speak, think, and respond strictly in English."
)

def respond(message, chat_history):
    # Format the chat history tracking layout for Gemma 3
    formatted_prompt = f"<|system|>\n{SYSTEM_PROMPT}<|end|>\n"
    
    for user_msg, ai_msg in chat_history:
        if user_msg:
            formatted_prompt += f"<|user|>\n{user_msg}<|end|>\n"
        if ai_msg:
            formatted_prompt += f"<|assistant|>\n{ai_msg}<|end|>\n"
            
    # Append the latest user input
    formatted_prompt += f"<|user|>\n{message}<|end|>\n<|assistant|>\n"
    
    # Generate the response completely inside the HF Cloud server
    response = llm(
        formatted_prompt,
        max_tokens=256,
        temperature=0.7,
        top_p=0.95,
        stop=["<|end|>", "###", "<|user|>"]
    )
    
    return response["choices"][0]["text"].strip()

# 4. Assemble the web browser layout interface
with ui.Blocks() as demo:
    ui.Markdown("# 🤖 Resin by Brody Systems")
    ui.Markdown("Running completely in the cloud. Casual, human, and lightweight.")
    ui.ChatInterface(fn=respond)

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