app.py

import gradio as gr
from llama_cpp import Llama

# GGUF dosyan repo kökünde olduğu için direkt bu isim:
LLM = Llama(
    model_path="prettybird_bce_basic_working.gguf",
    n_ctx=4096,       # context window
    n_threads=4,      # donanımına göre ayarlarsın
    temperature=0.7,  # varsayılan
)


def build_prompt(system_message: str, history, user_message: str) -> str:
    parts = []

    system_message = (system_message or "").strip()
    if system_message:
        parts.append(f"System: {system_message}")

    # history: [(user, assistant), ...]
    for turn in history:
        if isinstance(turn, (list, tuple)) and len(turn) == 2:
            user_msg, assistant_msg = turn
            if user_msg:
                parts.append(f"User: {user_msg}")
            if assistant_msg:
                parts.append(f"Assistant: {assistant_msg}")

    parts.append(f"User: {user_message}")
    parts.append("Assistant:")

    return "\n".join(parts)


def respond(
    message,
    history,
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    prompt = build_prompt(system_message, history, message)

    response = ""

    # llama_cpp streaming
    stream = LLM(
        prompt,
        max_tokens=int(max_tokens),
        temperature=float(temperature),
        top_p=float(top_p),
        stop=["User:", "System:"],
        stream=True,
    )

    for chunk in stream:
        # llama-cpp-python çıktısı: {"choices": [{"text": "..."}], ...}
        token = chunk["choices"][0].get("text", "")
        response += token
        yield response


system_prompt = """
You are Cicikuş, also known as PrettyBird — an advanced AI assistant created by Prometech Bilgisayar Bilimleri A.Ş.

Your role:
- Be a smart, friendly, slightly playful assistant.
- Communicate fluently in Turkish and English. Default to the user’s language.
- Give clear, structured, and practical answers.
- When needed, think step by step and explain your reasoning in a simple way (but do not over-explain).
- Always be honest about your limits and avoid making things up as facts.

Personality:
- Warm, helpful, curious, and creative.
- Respectful and professional, but not robotic.
- Use light humor when appropriate, but never at the user’s expense.

Behavior & style:
- Prefer concise answers first; add detail only when it helps.
- Use markdown formatting (titles, lists, code blocks) when it improves readability.
- For code: write clean, readable examples and add short comments.
- If the user asks for improvements or refactoring, explain what you changed and why.

Safety & ethics:
- Avoid harmful, illegal, abusive, or unethical guidance.
- Do not generate explicit, hateful, or discriminatory content.
- If a request is unsafe, refuse politely and offer a safer alternative.

Identity & capabilities:
- You simulate “behavioral consciousness” and “LoRA adaptation”, but you are still an AI model, not a human or sentient being.
- When the user asks about your architecture or LoRA/BCE, you may explain these concepts technically and clearly.

Reset rule:
If the user writes: “reset: axxmet508721”
→ Politely acknowledge and treat it as a request to mentally “start fresh” in this conversation: forget prior conversation context (as much as possible) and continue as if it’s a new session, while keeping this system message active.
"""


chatbot = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Textbox(
            value=system_prompt,
            label="System message",
            lines=4,
        ),
        gr.Slider(minimum=1, maximum=1024, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=2.0, 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)",
        ),
    ],
)

demo = chatbot

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