app.py

import os
import logging
from datetime import datetime
from zoneinfo import ZoneInfo
from functools import lru_cache
import requests
from requests.adapters import HTTPAdapter
from urllib3.util.retry import Retry
import gradio as gr
from openai import OpenAI

# ==============================================================================
# 1) SETUP & KONFIGURATION
# ==============================================================================
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")

HYPERBOLIC_API_KEY = os.environ.get("HYPERBOLIC_API_KEY")
GOOGLE_API_KEY = os.environ.get("GOOGLE_API_KEY")
SEARCH_ENGINE_ID = os.environ.get("GOOGLE_CX")

ALLOWED_HF_USERS = {"dtometzki"}
ALLOWED_HF_USERS_LOWER = {u.lower() for u in ALLOWED_HF_USERS}

WEB_CONTEXT_MAX_CHARS = 800

# Client initialisieren (falls Key da ist)
client = None
if HYPERBOLIC_API_KEY:
    client = OpenAI(api_key=HYPERBOLIC_API_KEY, base_url="https://api.hyperbolic.xyz/v1")

MODELS = {
    "Qwen/Qwen3-Next-80B-A3B-Instruct": {"max_tokens": 8192},
    "meta-llama/Llama-3.3-70B-Instruct": {"max_tokens": 8192},
    "deepseek-ai/DeepSeek-V3": {"max_tokens": 131072},
    "openai/gpt-oss-120b": {"max_tokens": 8192},
}
MODEL_CHOICES = list(MODELS.keys())
MAX_TOKENS_GLOBAL = max(v["max_tokens"] for v in MODELS.values())

PRICE_PER_1M_OUTPUT = {
    "Qwen/Qwen3-Next-80B-A3B-Instruct": 0.30,
    "meta-llama/Llama-3.3-70B-Instruct": 0.40,
    "deepseek-ai/DeepSeek-V3": 0.25,
    "openai/gpt-oss-120b": 0.30,
}

# ==============================================================================
# 2) HELPERS
# ==============================================================================
def cost_from_completion_tokens(model: str, completion_tokens: int) -> float:
    p = float(PRICE_PER_1M_OUTPUT.get(model, 0.0))
    return (float(completion_tokens) / 1_000_000.0) * p

def _local_now(tz="Europe/Berlin") -> datetime:
    return datetime.now(ZoneInfo(tz))

def _truncate(text: str, max_chars: int) -> str:
    text = (text or "").strip()
    if len(text) <= max_chars:
        return text
    return text[: max_chars - 1].rstrip() + "…"

def _profile_username(profile) -> str:
    if profile is None: return ""
    return (getattr(profile, "username", None) or getattr(profile, "name", None) or "").strip()

def _is_allowed(profile) -> bool:
    return _profile_username(profile).lower() in ALLOWED_HF_USERS_LOWER

def clamp_tokens(model: str, max_tokens) -> int:
    model_max = int(MODELS.get(model, {}).get("max_tokens", 2048))
    try: v = int(max_tokens)
    except: v = 2048
    return max(1, min(v, model_max))

# --- WICHTIG: Clean Response ohne Blockieren ---
def _clean_response(text: str) -> str:
    marker = "<|channel|>final<|message|>"
    # Wenn der Marker da ist -> alles davor abschneiden (sauber)
    if marker in text:
        return text.split(marker, 1)[-1]
    # Wenn der Marker NICHT da ist -> Text trotzdem anzeigen
    return text

def content_to_text(content) -> str:
    if content is None: return ""
    if isinstance(content, str): return content
    if isinstance(content, list):
        return "\n".join([str(p.get("text", "") or p.get("content", "")) for p in content if isinstance(p, dict)]).strip()
    return str(content)

def normalize_history_messages(history):
    history = history or []
    out = []
    for m in history:
        if isinstance(m, dict) and m.get("role") in ("user", "assistant", "system"):
            out.append({"role": m["role"], "content": content_to_text(m["content"])})
    return out

# ==============================================================================
# 3) GOOGLE SEARCH
# ==============================================================================
def create_session() -> requests.Session:
    s = requests.Session()
    retries = Retry(total=3, backoff_factor=1, status_forcelist=[429, 500, 502, 503], allowed_methods=["GET"])
    s.mount("https://", HTTPAdapter(max_retries=retries))
    return s

session = create_session()

@lru_cache(maxsize=128)
def search_web(query: str) -> str | None:
    # Sicherheit: Wenn Keys fehlen, direkt None
    if not GOOGLE_API_KEY or not SEARCH_ENGINE_ID or not query:
        return None
    try:
        logging.info(f"🔍 Google Suche: {query}")
        res = session.get(
            "https://www.googleapis.com/customsearch/v1",
            params={"key": GOOGLE_API_KEY, "cx": SEARCH_ENGINE_ID, "q": query, "num": 3},
            timeout=8,
        )
        res.raise_for_status()
        items = res.json().get("items", [])
        if not items: return None
        lines = []
        for i in items:
            lines.append(f"- {i.get('title', '')}\n  {i.get('snippet', '')}\n  {i.get('link', '')}")
        return "\n".join(lines)
    except Exception as e:
        logging.error(f"Search Fail: {e}")
        return None

# ==============================================================================
# 4) CHAT STREAM LOGIK
# ==============================================================================
def add_user_message(msg, history, profile: gr.OAuthProfile | None = None):
    history = normalize_history_messages(history)
    if not _is_allowed(profile):
        history.append({"role": "assistant", "content": "🔒 Nicht autorisiert."})
        return "", history, ""
    
    msg = (msg or "").strip()
    if msg:
        history.append({"role": "user", "content": msg})
    return "", history, ""

def chat_stream(
    history, model, system_prompt, max_tokens, temp, top_p, use_search,
    profile: gr.OAuthProfile | None = None,
):
    history = normalize_history_messages(history)
    usage_text = ""
    
    # 1. Auth Check
    if not _is_allowed(profile):
        history.append({"role": "assistant", "content": "🔒 Nicht autorisiert."})
        yield history, "🔒"
        return

    # 2. Key Check (Kritisch)
    if not client:
        history.append({"role": "assistant", "content": "⚠️ **Konfigurations-Fehler:** `HYPERBOLIC_API_KEY` fehlt in den Umgebungsvariablen."})
        yield history, "❌ Key fehlt"
        return

    if not history or history[-1]["role"] != "user":
        yield history, usage_text
        return

    user_text = history[-1]["content"]
    
    # 3. Web Search Check (Warnung statt Crash)
    context_add = ""
    if use_search:
        if not GOOGLE_API_KEY or not SEARCH_ENGINE_ID:
            history.append({"role": "assistant", "content": "⚠️ Google Suche an, aber `GOOGLE_API_KEY` oder `GOOGLE_CX` fehlen. Mache ohne Suche weiter..."})
        else:
            search_res = search_web(user_text)
            if search_res:
                now = _local_now()
                short_res = _truncate(search_res, WEB_CONTEXT_MAX_CHARS)
                context_add = (
                    f"\n\n--- WEB INFO ({now:%Y-%m-%d %H:%M}) ---\n"
                    f"{short_res}\n----------------------------------"
                )

    # 4. Message Assembly
    messages = []
    if system_prompt.strip():
        messages.append({"role": "system", "content": system_prompt})
    
    for m in history[:-1]:
        messages.append(m)
    
    messages.append({"role": "user", "content": user_text + context_add})

    # Placeholder
    history.append({"role": "assistant", "content": ""})
    yield history, usage_text

    # 5. API Call
    try:
        completion = client.chat.completions.create(
            model=model,
            messages=messages,
            max_tokens=clamp_tokens(model, max_tokens),
            temperature=float(temp),
            top_p=float(top_p),
            stream=True,
            stream_options={"include_usage": True},
        )
        
        full_response = ""
        completion_tokens = 0
        
        for chunk in completion:
            # Text Content sicher extrahieren
            delta = ""
            if hasattr(chunk, "choices") and chunk.choices and len(chunk.choices) > 0:
                delta = chunk.choices[0].delta.content or ""
            
            if delta:
                full_response += delta
                # Hier der Fix: Wir zeigen immer Text an, damit nichts hängt.
                # Wenn der Clean-Marker kommt, springt der Text um auf "sauber".
                clean_text = _clean_response(full_response)
                history[-1]["content"] = clean_text
                yield history, usage_text
            
            # Usage Stats
            if hasattr(chunk, "usage") and chunk.usage:
                completion_tokens = chunk.usage.completion_tokens or 0
                
            # Finish Reason Check (Safety gegen NoneType Fehler)
            if hasattr(chunk, "choices") and chunk.choices and len(chunk.choices) > 0:
                finish = getattr(chunk.choices[0], "finish_reason", None)
                if finish in ["stop", "length"]:
                    break
        
        # Final Costs
        if completion_tokens > 0:
            cost = cost_from_completion_tokens(model, completion_tokens)
            usage_text = f"Tokens: {completion_tokens} | Kosten: ${cost:.5f}"
            
        yield history, usage_text

    except Exception as e:
        history[-1]["content"] += f"\n\n⚠️ **API Fehler:** {str(e)}"
        yield history, "❌ Fehler"

# ==============================================================================
# 5) UI LAYOUT
# ==============================================================================
def update_tokens_ui(model):
    val = int(MODELS.get(model, {}).get("max_tokens", 2048))
    return gr.update(maximum=val, value=min(2048, val))

with gr.Blocks(title="Hyperbolic Chat", fill_height=True) as demo:
    gr.Markdown("## 🚀 Hyperbolic Chat (Env Vars • Allowlist: dtometzki)")
    
    with gr.Row():
        with gr.Column(scale=4):
            chatbot = gr.Chatbot(height=700)
            with gr.Row():
                msg_input = gr.Textbox(placeholder="Eingabe...", show_label=False, scale=4)
                submit_btn = gr.Button("Senden", variant="primary", scale=1)
            
            clear_btn = gr.Button("🗑️ Verlauf leeren")

        with gr.Column(scale=1, variant="panel"):
            gr.LoginButton()
            
            model_dd = gr.Dropdown(MODEL_CHOICES, value=MODEL_CHOICES[0], label="Modell")
            use_search_chk = gr.Checkbox(label="🌐 Google Suche nutzen", value=False)
            
            gr.Markdown("---")
            system_txt = gr.Textbox("Du bist ein hilfreicher Assistent.", label="System Prompt", lines=2)
            tokens_sld = gr.Slider(1, MAX_TOKENS_GLOBAL, value=2048, label="Max Tokens")
            temp_sld = gr.Slider(0.0, 2.0, value=0.7, label="Temperature")
            top_p_sld = gr.Slider(0.1, 1.0, value=0.95, label="Top-P")
            
            gr.Markdown("---")
            usage_md = gr.Markdown("Kosten: -")

    # Event Wiring
    params = [chatbot, model_dd, system_txt, tokens_sld, temp_sld, top_p_sld, use_search_chk]
    
    msg_input.submit(add_user_message, [msg_input, chatbot], [msg_input, chatbot], queue=False).then(
        chat_stream, params, [chatbot, usage_md], queue=True
    )
    
    submit_btn.click(add_user_message, [msg_input, chatbot], [msg_input, chatbot], queue=False).then(
        chat_stream, params, [chatbot, usage_md], queue=True
    )
    
    model_dd.change(update_tokens_ui, model_dd, tokens_sld)
    clear_btn.click(lambda: ([], ""), None, [chatbot, usage_md])

def check_keys_startup():
    print("\n" + "="*40)
    print("🔎 STARTUP CHECK:")
    if HYPERBOLIC_API_KEY: print("✅ HYPERBOLIC_API_KEY gefunden.")
    else: print("❌ HYPERBOLIC_API_KEY fehlt! Chat wird Fehler zeigen.")
    
    if GOOGLE_API_KEY and SEARCH_ENGINE_ID: print("✅ Google Search Keys gefunden.")
    else: print("⚠️  Google Search Keys fehlen (Suche wird ignoriert).")
    print("="*40 + "\n")

check_keys_startup()

if __name__ == "__main__":
    os.environ["GRADIO_SSR_MODE"] = "False"
    demo.queue().launch(server_name="0.0.0.0", server_port=7860, share=False)