app.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

import os, glob, textwrap
from pathlib import Path

import gradio as gr
from huggingface_hub import snapshot_download
from llama_cpp import Llama
import requests
from bs4 import BeautifulSoup

# ===== Modelo (GGUF) =====
MODEL_REPO    = os.getenv("MODEL_REPO", "Qwen/Qwen2.5-7B-Instruct-GGUF")
# Para CPU Basic puedes bajar a q3_k_m si te falta RAM
MODEL_PATTERN = os.getenv("MODEL_PATTERN", "qwen2.5-7b-instruct-q4_k_m-*.gguf")

LOCAL_DIR = Path("models"); LOCAL_DIR.mkdir(parents=True, exist_ok=True)
print(f"[Boot] Descargando {MODEL_REPO} patrón {MODEL_PATTERN} ...")
snapshot_dir = snapshot_download(repo_id=MODEL_REPO, local_dir=str(LOCAL_DIR),
                                 allow_patterns=[MODEL_PATTERN])
candidates = sorted(glob.glob(str(Path(snapshot_dir) / MODEL_PATTERN)))
if not candidates:
    raise FileNotFoundError(f"No hay shards para {MODEL_PATTERN} en {snapshot_dir}")
MODEL_PATH = candidates[0]
print(f"[Boot] Usando shard: {MODEL_PATH}")

N_THREADS = max(1, (os.cpu_count() or 2) - 1)

llm = Llama(
    model_path=MODEL_PATH,
    n_ctx=4096,
    n_threads=N_THREADS,
    n_batch=256,
    n_gpu_layers=0,
    verbose=False,
)

SYSTEM_DEFAULT = textwrap.dedent("""\
Eres Astrohunters-Guide, un asistente en español.
- Respondes con precisión y sin inventar datos.
- Sabes explicar resultados de exoplanetas (período, duración, profundidad, SNR, radio).
- Si te paso una URL, lees su contenido y lo usas como contexto.
""")

def fetch_url_text(url: str, max_chars: int = 6000) -> str:
    try:
        r = requests.get(url, timeout=15)
        r.raise_for_status()
        soup = BeautifulSoup(r.text, "html.parser")
        for t in soup(["script", "style", "noscript"]): t.decompose()
        txt = " ".join(soup.get_text(separator=" ").split())
        return txt[:max_chars]
    except Exception as e:
        return f"[No se pudo cargar {url}: {e}]"

def run_llm(messages, temperature=0.6, top_p=0.95, max_tokens=768):
    out = llm.create_chat_completion(
        messages=messages,
        temperature=temperature,
        top_p=top_p,
        max_tokens=max_tokens,
        stream=False,
    )
    return out["choices"][0]["message"]["content"].strip()

# ====== Funciones API ======
def api_run_predict(prompt: str, system: str = "") -> str:
    messages = [
        {"role": "system", "content": system or SYSTEM_DEFAULT},
        {"role": "user", "content": prompt},
    ]
    return run_llm(messages, max_tokens=512)

def api_run_predict_with_url(prompt: str, url: str = "", system: str = "") -> str:
    web_ctx = fetch_url_text(url) if url else ""
    user_msg = prompt if not web_ctx else f"{prompt}\n\n[CONTEXTO_WEB]\n{web_ctx}"
    messages = [
        {"role": "system", "content": system or SYSTEM_DEFAULT},
        {"role": "user", "content": user_msg},
    ]
    return run_llm(messages, max_tokens=700)

# ===== UI de chat =====
with gr.Blocks(title="Astrohunters LLM (Qwen2.5 7B)") as chat_ui:
    gr.Markdown("## 🛰️ Astrohunters LLM (Qwen2.5 7B Instruct, GGUF — CPU Basic)")
    with gr.Row():
        with gr.Column(scale=3):
            chat = gr.Chatbot(height=420, type="tuples")
            with gr.Row():
                txt = gr.Textbox(placeholder="Escribe tu pregunta...", scale=4)
                btn = gr.Button("Enviar", scale=1, variant="primary")
        with gr.Column(scale=2):
            system_tb = gr.Textbox(label="System prompt", value=SYSTEM_DEFAULT, lines=10)
            url_tb = gr.Textbox(label="URL (opcional): Cargar contenido web", placeholder="https://...")

    def chat_infer(history, system_prompt, user, url_to_load):
        web_ctx = fetch_url_text(url_to_load.strip()) if url_to_load and url_to_load.strip() else ""
        messages = [{"role": "system", "content": system_prompt or SYSTEM_DEFAULT}]
        for u, a in history:
            if u: messages.append({"role": "user", "content": u})
            if a: messages.append({"role": "assistant", "content": a})
        user_msg = user or ""
        if web_ctx:
            user_msg = f"{user_msg}\n\n[CONTEXTO_WEB]\n{web_ctx}"
        messages.append({"role": "user", "content": user_msg})
        reply = run_llm(messages, max_tokens=700)
        history.append((user, reply))
        return history, ""

    btn.click(chat_infer, inputs=[chat, system_tb, txt, url_tb], outputs=[chat, txt])
    txt.submit(chat_infer, inputs=[chat, system_tb, txt, url_tb], outputs=[chat, txt])

# ====== APIs nombradas (fuera del Blocks) ======
api1 = gr.Interface(
    fn=api_run_predict,
    inputs=[gr.Textbox(label="prompt"), gr.Textbox(label="system")],
    outputs=gr.Textbox(label="reply"),
    api_name="run_predict",
)

api2 = gr.Interface(
    fn=api_run_predict_with_url,
    inputs=[gr.Textbox(label="prompt"), gr.Textbox(label="url"), gr.Textbox(label="system")],
    outputs=gr.Textbox(label="reply"),
    api_name="run_predict_with_url",
)

# Unimos todo en un solo demo para que Gradio registre las rutas
demo = gr.TabbedInterface(
    [chat_ui, api1, api2],
    tab_names=["Chat", "API: run_predict", "API: run_predict_with_url"],
)

if __name__ == "__main__":
    demo.queue(max_size=16).launch(server_name="0.0.0.0", server_port=7860)