api.py

from __future__ import annotations
from fastapi import FastAPI, UploadFile, File, Form, BackgroundTasks, HTTPException
from fastapi import Body
from fastapi.responses import JSONResponse, FileResponse
from fastapi.middleware.cors import CORSMiddleware
from pathlib import Path
import shutil
import uvicorn
import json
import uuid
from datetime import datetime
from typing import Dict
from enum import Enum
import os

from video_processing import process_video_pipeline
from audio_tools import process_audio_for_video, extract_audio_ffmpeg, embed_voice_segments
from casting_loader import ensure_chroma, build_faces_index, build_voices_index
from narration_system import NarrationSystem
from llm_router import load_yaml, LLMRouter
from character_detection import detect_characters_from_video

from pipelines.audiodescription import generate as ad_generate

app = FastAPI(title="Veureu Engine API", version="0.2.0")
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

ROOT = Path("/tmp/veureu")
ROOT.mkdir(parents=True, exist_ok=True)
TEMP_ROOT = Path("/tmp/temp")
TEMP_ROOT.mkdir(parents=True, exist_ok=True)
VIDEOS_ROOT = Path("/tmp/data/videos")
VIDEOS_ROOT.mkdir(parents=True, exist_ok=True)
IDENTITIES_ROOT = Path("/tmp/characters")
IDENTITIES_ROOT.mkdir(parents=True, exist_ok=True)

# Sistema de jobs asíncronos
class JobStatus(str, Enum):
    QUEUED = "queued"
    PROCESSING = "processing"
    DONE = "done"
    FAILED = "failed"

jobs: Dict[str, dict] = {}

def describe_image_with_svision(image_path: str, is_face: bool = True) -> tuple[str, str]:
    """
    Llama al space svision para describir una imagen (usado en generación de AD).
    
    Args:
        image_path: Ruta absoluta a la imagen
        is_face: True si es una cara, False si es una escena
    
    Returns:
        tuple (descripción_completa, nombre_abreviado)
    """
    try:
        from pathlib import Path as _P
        import yaml
        from llm_router import LLMRouter
        
        # Cargar configuración
        config_path = _P(__file__).parent / "config.yaml"
        if not config_path.exists():
            print(f"[svision] Config no encontrado: {config_path}")
            return ("", "")
        
        with open(config_path, 'r', encoding='utf-8') as f:
            cfg = yaml.safe_load(f) or {}
        
        router = LLMRouter(cfg)
        
        # Contexto diferente para caras vs escenas
        if is_face:
            context = {
                "task": "describe_person",
                "instructions": "Descriu la persona en la imatge. Inclou: edat aproximada (jove/adult), gènere, característiques físiques notables (ulleres, barba, bigoti, etc.), expressió i vestimenta."
            }
        else:
            context = {
                "task": "describe_scene",
                "instructions": "Descriu l'escena en la imatge. Inclou: tipus de localització (interior/exterior), elements principals, ambient, il·luminació."
            }
        
        # Llamar a svision
        descriptions = router.vision_describe([str(image_path)], context=context, model="salamandra-vision")
        full_description = descriptions[0] if descriptions else ""
        
        if not full_description:
            return ("", "")
        
        print(f"[svision] Descripció generada: {full_description[:100]}...")
        
        return (full_description, "")
        
    except Exception as e:
        print(f"[svision] Error al descriure imatge: {e}")
        import traceback
        traceback.print_exc()
        return ("", "")

def normalize_face_lighting(image):
    """
    Normaliza el brillo de una imagen de cara usando técnicas combinadas:
    1. CLAHE para ecualización adaptativa
    2. Normalización de rango para homogeneizar brillo general
    
    Esto reduce el impacto de diferentes condiciones de iluminación en los embeddings
    y en la visualización de las imágenes.
    
    Args:
        image: Imagen BGR (OpenCV format)
    
    Returns:
        Imagen normalizada en el mismo formato
    """
    import cv2
    import numpy as np
    
    # Paso 1: Convertir a LAB color space (más robusto para iluminación)
    lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB)
    l, a, b = cv2.split(lab)
    
    # Paso 2: Aplicar CLAHE (Contrast Limited Adaptive Histogram Equalization) al canal L
    # Usar clipLimit más alto para normalización más agresiva
    clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8, 8))
    l_clahe = clahe.apply(l)
    
    # Paso 3: Normalizar el rango del canal L para asegurar distribución uniforme
    # Esto garantiza que todas las imágenes tengan un rango de brillo similar
    l_min, l_max = l_clahe.min(), l_clahe.max()
    if l_max > l_min:
        # Estirar el histograma al rango completo [0, 255]
        l_normalized = ((l_clahe - l_min) * 255.0 / (l_max - l_min)).astype(np.uint8)
    else:
        l_normalized = l_clahe
    
    # Paso 4: Aplicar suavizado suave para reducir ruido introducido por la normalización
    l_normalized = cv2.GaussianBlur(l_normalized, (3, 3), 0)
    
    # Recombinar canales
    lab_normalized = cv2.merge([l_normalized, a, b])
    
    # Convertir de vuelta a BGR
    normalized = cv2.cvtColor(lab_normalized, cv2.COLOR_LAB2BGR)
    return normalized

def hierarchical_cluster_with_min_size(X, max_groups: int, min_cluster_size: int):
    """
    Clustering jerárquico aglomerativo con selección óptima del número de clusters.
    Selecciona automáticamente el mejor número de clusters (hasta max_groups) usando silhouette score.
    Filtra clusters con menos de min_cluster_size muestras (marcados como -1/ruido).
    
    Args:
        X: Array de embeddings (N, D)
        max_groups: Número máximo de clusters a formar
        min_cluster_size: Tamaño mínimo de cluster válido
        
    Returns:
        Array de labels (N,) donde -1 indica ruido
    """
    import numpy as np
    from scipy.cluster.hierarchy import linkage, fcluster
    from sklearn.metrics import silhouette_score
    from collections import Counter
    
    if len(X) == 0:
        return np.array([])
    
    if len(X) < min_cluster_size:
        # Si hay menos muestras que el mínimo, todo es ruido
        return np.full(len(X), -1, dtype=int)
    
    # Linkage usando distancia euclidiana con método 'ward'
    Z = linkage(X, method='ward', metric='euclidean')
    
    # Encontrar el número óptimo de clusters usando silhouette score
    best_n_clusters = 2
    best_score = -1
    
    # Probar diferentes números de clusters (de 2 a max_groups)
    max_to_try = min(max_groups, len(X) - 1)  # No puede haber más clusters que muestras
    
    if max_to_try >= 2:
        for n_clusters in range(2, max_to_try + 1):
            trial_labels = fcluster(Z, t=n_clusters, criterion='maxclust') - 1
            
            # Calcular cuántos clusters válidos tendríamos después del filtrado
            trial_counts = Counter(trial_labels)
            valid_clusters = sum(1 for count in trial_counts.values() if count >= min_cluster_size)
            
            # Solo evaluar si hay al menos 2 clusters válidos
            if valid_clusters >= 2:
                try:
                    score = silhouette_score(X, trial_labels, metric='euclidean')
                    # Penalizar ligeramente configuraciones con muchos clusters para evitar overfitting
                    adjusted_score = score - (n_clusters * 0.01)
                    
                    if adjusted_score > best_score:
                        best_score = adjusted_score
                        best_n_clusters = n_clusters
                except:
                    pass  # Si falla el cálculo, ignorar esta configuración
    
    # Usar el número óptimo de clusters encontrado
    print(f"Clustering óptimo: {best_n_clusters} clusters (de máximo {max_groups}), silhouette score: {best_score:.3f}")
    labels = fcluster(Z, t=best_n_clusters, criterion='maxclust')
    
    # fcluster devuelve labels 1-indexed, convertir a 0-indexed
    labels = labels - 1
    
    # Filtrar clusters pequeños
    label_counts = Counter(labels)
    filtered_labels = []
    for lbl in labels:
        if label_counts[lbl] >= min_cluster_size:
            filtered_labels.append(lbl)
        else:
            filtered_labels.append(-1)  # Ruido
    
    return np.array(filtered_labels, dtype=int)

@app.get("/")
def root():
    return {"ok": True, "service": "veureu-engine"}

@app.post("/process_video")
async def process_video(
    video_file: UploadFile = File(...),
    config_path: str = Form("config.yaml"),
    out_root: str = Form("results"),
    db_dir: str = Form("chroma_db"),
):
    tmp_video = ROOT / video_file.filename
    with tmp_video.open("wb") as f:
        shutil.copyfileobj(video_file.file, f)
    result = process_video_pipeline(str(tmp_video), config_path=config_path, out_root=out_root, db_dir=db_dir)
    return JSONResponse(result)

@app.post("/create_initial_casting")
async def create_initial_casting(
    background_tasks: BackgroundTasks,
    video: UploadFile = File(...),
    max_groups: int = Form(default=3),
    min_cluster_size: int = Form(default=3),
    voice_max_groups: int = Form(default=3),
    voice_min_cluster_size: int = Form(default=3),
    max_frames: int = Form(default=100),
):
    """
    Crea un job para procesar el vídeo de forma asíncrona usando clustering jerárquico.
    Devuelve un job_id inmediatamente.
    """
    # Guardar vídeo en carpeta de datos
    video_name = Path(video.filename).stem
    dst_video = VIDEOS_ROOT / f"{video_name}.mp4"
    with dst_video.open("wb") as f:
        shutil.copyfileobj(video.file, f)

    # Crear job_id único
    job_id = str(uuid.uuid4())
    
    # Inicializar el job
    jobs[job_id] = {
        "id": job_id,
        "status": JobStatus.QUEUED,
        "video_path": str(dst_video),
        "video_name": video_name,
        "max_groups": int(max_groups),
        "min_cluster_size": int(min_cluster_size),
        "voice_max_groups": int(voice_max_groups),
        "voice_min_cluster_size": int(voice_min_cluster_size),
        "max_frames": int(max_frames),
        "created_at": datetime.now().isoformat(),
        "results": None,
        "error": None
    }
    
    print(f"[{job_id}] Job creado para vídeo: {video_name}")
    
    # Iniciar procesamiento en background
    background_tasks.add_task(process_video_job, job_id)
    
    # Devolver job_id inmediatamente
    return {"job_id": job_id}

@app.get("/jobs/{job_id}/status")
def get_job_status(job_id: str):
    """
    Devuelve el estado actual de un job.
    El UI hace polling de este endpoint cada 5 segundos.
    """
    if job_id not in jobs:
        raise HTTPException(status_code=404, detail="Job not found")
    
    job = jobs[job_id]
    
    # Normalizar el estado a string
    status_value = job["status"].value if isinstance(job["status"], JobStatus) else str(job["status"])
    response = {"status": status_value}

    # Incluir resultados si existen (evita condiciones de carrera)
    if job.get("results") is not None:
        response["results"] = job["results"]

    # Incluir error si existe
    if job.get("error"):
        response["error"] = job["error"]
    
    return response

@app.get("/files/{video_name}/{char_id}/{filename}")
def serve_character_file(video_name: str, char_id: str, filename: str):
    """
    Sirve archivos estáticos de personajes (imágenes).
    Ejemplo: /files/dif_catala_1/char1/representative.jpg
    """
    # Las caras se guardan en /tmp/temp/<video>/characters/<char_id>/<filename>
    file_path = TEMP_ROOT / video_name / "characters" / char_id / filename
    
    if not file_path.exists():
        raise HTTPException(status_code=404, detail="File not found")
    
    return FileResponse(file_path)

@app.get("/audio/{video_name}/{filename}")
def serve_audio_file(video_name: str, filename: str):
    file_path = TEMP_ROOT / video_name / "clips" / filename
    if not file_path.exists():
        raise HTTPException(status_code=404, detail="File not found")
    return FileResponse(file_path)

def process_video_job(job_id: str):
    """
    Procesa el vídeo de forma asíncrona.
    Esta función se ejecuta en background.
    """
    try:
        job = jobs[job_id]
        print(f"[{job_id}] Iniciando procesamiento...")
        
        # Cambiar estado a processing
        job["status"] = JobStatus.PROCESSING
        
        video_path = job["video_path"]
        video_name = job["video_name"]
        max_groups = int(job.get("max_groups", 5))
        min_cluster_size = int(job.get("min_cluster_size", 3))
        v_max_groups = int(job.get("voice_max_groups", 5))
        v_min_cluster = int(job.get("voice_min_cluster_size", 3))
        
        # Crear estructura de carpetas
        base = TEMP_ROOT / video_name
        base.mkdir(parents=True, exist_ok=True)
        
        print(f"[{job_id}] Directorio base: {base}")
        
        # Detección de caras y embeddings (CPU), alineado con 'originales'
        try:
            print(f"[{job_id}] Iniciando detección de personajes (CPU, originales)...")
            print(f"[{job_id}] *** Normalización de brillo ACTIVADA ***")
            print(f"[{job_id}]   - CLAHE adaptativo (clipLimit=3.0)")
            print(f"[{job_id}]   - Estiramiento de histograma")
            print(f"[{job_id}]   - Suavizado Gaussiano")
            print(f"[{job_id}]   Esto homogeneizará el brillo de todas las caras detectadas")
            import cv2
            import numpy as np
            try:
                import face_recognition  # CPU
                _use_fr = True
                print(f"[{job_id}] face_recognition disponible: CPU")
            except Exception:
                face_recognition = None  # type: ignore
                _use_fr = False
                print(f"[{job_id}] face_recognition no disponible. Intentando DeepFace fallback.")
                try:
                    from deepface import DeepFace  # type: ignore
                except Exception:
                    DeepFace = None  # type: ignore
            
            cap = cv2.VideoCapture(video_path)
            if not cap.isOpened():
                raise RuntimeError("No se pudo abrir el vídeo para extracción de caras")
            fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
            total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT) or 0)
            max_samples = job.get("max_frames", 100)
            # Índices de frames equiespaciados
            if total_frames > 0:
                frame_indices = sorted(set(np.linspace(0, max(0, total_frames - 1), num=min(max_samples, max(1, total_frames)), dtype=int).tolist()))
            else:
                frame_indices = []
            print(f"[{job_id}] Total frames: {total_frames}, FPS: {fps:.2f}, Muestreando {len(frame_indices)} frames equiespaciados (máx {max_samples})")

            # Salidas
            faces_root = base / "faces_raw"
            faces_root.mkdir(parents=True, exist_ok=True)
            embeddings: list[list[float]] = []
            crops_meta: list[dict] = []

            saved_count = 0
            frames_processed = 0
            frames_with_faces = 0
            for frame_idx in frame_indices:
                cap.set(cv2.CAP_PROP_POS_FRAMES, int(frame_idx))
                ret2, frame = cap.read()
                if not ret2:
                    continue
                frames_processed += 1
                # Normalizar iluminación antes de procesar
                frame_normalized = normalize_face_lighting(frame)
                rgb = cv2.cvtColor(frame_normalized, cv2.COLOR_BGR2RGB)

                if _use_fr and face_recognition is not None:
                    boxes = face_recognition.face_locations(rgb, model="hog")  # CPU HOG
                    encs = face_recognition.face_encodings(rgb, boxes)
                    if boxes:
                        frames_with_faces += 1
                        print(f"[{job_id}] Frame {frame_idx}: {len(boxes)} cara(s) detectada(s) con face_recognition")
                    for (top, right, bottom, left), e in zip(boxes, encs):
                        crop = frame_normalized[top:bottom, left:right]
                        if crop.size == 0:
                            continue
                        fn = f"face_{frame_idx:06d}_{saved_count:03d}.jpg"
                        cv2.imwrite(str(faces_root / fn), crop)
                        # Normalizar embedding
                        e = np.array(e, dtype=float)
                        e = e / (np.linalg.norm(e) + 1e-9)
                        embeddings.append(e.astype(float).tolist())
                        crops_meta.append({
                            "file": fn,
                            "frame": frame_idx,
                            "box": [int(top), int(right), int(bottom), int(left)],
                        })
                        saved_count += 1
                else:
                    # DeepFace fallback con detección de bounding boxes vía Haar Cascade (OpenCV)
                    if DeepFace is None:
                        pass
                    else:
                        try:
                            gray = cv2.cvtColor(frame_normalized, cv2.COLOR_BGR2GRAY)
                            try:
                                haar_path = getattr(cv2.data, 'haarcascades', None) or ''
                                face_cascade = cv2.CascadeClassifier(os.path.join(haar_path, 'haarcascade_frontalface_default.xml'))
                            except Exception:
                                face_cascade = None
                            boxes_haar = []
                            if face_cascade is not None and not face_cascade.empty():
                                # Parámetros más estrictos para evitar falsos positivos
                                faces_haar = face_cascade.detectMultiScale(gray, scaleFactor=1.08, minNeighbors=5, minSize=(50, 50))
                                for (x, y, w, h) in faces_haar:
                                    top, left, bottom, right = max(0, y), max(0, x), min(frame.shape[0], y+h), min(frame.shape[1], x+w)
                                    boxes_haar.append((top, right, bottom, left))
                            
                            # Si Haar no detecta nada, intentar con DeepFace directamente
                            if not boxes_haar:
                                try:
                                    tmp_detect = faces_root / f"detect_{frame_idx:06d}.jpg"
                                    cv2.imwrite(str(tmp_detect), frame_normalized)
                                    detect_result = DeepFace.extract_faces(img_path=str(tmp_detect), detector_backend='opencv', enforce_detection=False)
                                    for det in detect_result:
                                        facial_area = det.get('facial_area', {})
                                        if facial_area:
                                            x, y, w, h = facial_area.get('x', 0), facial_area.get('y', 0), facial_area.get('w', 0), facial_area.get('h', 0)
                                            # Validar que es un bbox real, no el frame completo
                                            # Si el bbox es prácticamente el frame completo, descartarlo
                                            is_full_frame = (x <= 5 and y <= 5 and w >= frame.shape[1] - 10 and h >= frame.shape[0] - 10)
                                            # Bbox mínimo de 50x50 para filtrar falsos positivos pequeños
                                            if w > 50 and h > 50 and not is_full_frame:
                                                top, left, bottom, right = max(0, y), max(0, x), min(frame.shape[0], y+h), min(frame.shape[1], x+w)
                                                boxes_haar.append((top, right, bottom, left))
                                    tmp_detect.unlink(missing_ok=True)
                                except Exception as _e_detect:
                                    print(f"[{job_id}] Frame {frame_idx}: DeepFace extract_faces error: {_e_detect}")
                            
                            if boxes_haar:
                                frames_with_faces += 1
                                print(f"[{job_id}] Frame {frame_idx}: {len(boxes_haar)} cara(s) detectada(s) con Haar/DeepFace")
                            
                            for (top, right, bottom, left) in boxes_haar:
                                crop = frame_normalized[top:bottom, left:right]
                                if crop.size == 0:
                                    continue
                                fn = f"face_{frame_idx:06d}_{saved_count:03d}.jpg"
                                crop_path = faces_root / fn
                                cv2.imwrite(str(crop_path), crop)
                                reps = DeepFace.represent(img_path=str(crop_path), model_name="Facenet512", enforce_detection=False)
                                for r in (reps or []):
                                    emb = r.get("embedding") if isinstance(r, dict) else r
                                    if emb is None:
                                        continue
                                    emb = np.array(emb, dtype=float)
                                    emb = emb / (np.linalg.norm(emb) + 1e-9)
                                    embeddings.append(emb.astype(float).tolist())
                                    crops_meta.append({
                                        "file": fn,
                                        "frame": frame_idx,
                                        "box": [int(top), int(right), int(bottom), int(left)],
                                    })
                                    saved_count += 1
                        except Exception as _e_df:
                            print(f"[{job_id}] DeepFace fallback error: {_e_df}")
            cap.release()

            print(f"[{job_id}] ✓ Frames procesados: {frames_processed}/{len(frame_indices)}")
            print(f"[{job_id}] ✓ Frames con caras: {frames_with_faces}")
            print(f"[{job_id}] ✓ Caras detectadas (embeddings): {len(embeddings)}")

            # Clustering jerárquico de caras
            if embeddings:
                Xf = np.array(embeddings)
                labels = hierarchical_cluster_with_min_size(Xf, max_groups, min_cluster_size).tolist()
                print(f"[{job_id}] Clustering jerárquico de caras: {len(set([l for l in labels if l >= 0]))} clusters")
            else:
                labels = []

            # Construir carpetas por clúster y representative
            characters = []
            cluster_map: dict[int, list[int]] = {}
            for i, lbl in enumerate(labels):
                if isinstance(lbl, int) and lbl >= 0:
                    cluster_map.setdefault(lbl, []).append(i)

            chars_dir = base / "characters"
            chars_dir.mkdir(parents=True, exist_ok=True)
            import shutil as _sh
            for ci, idxs in sorted(cluster_map.items(), key=lambda x: x[0]):
                char_id = f"char_{ci:02d}"
                out_dir = chars_dir / char_id
                out_dir.mkdir(parents=True, exist_ok=True)
                files = []
                for k, j in enumerate(idxs[:24]):  # limitar a 24
                    fname = crops_meta[j]["file"]
                    src = faces_root / fname
                    dst = out_dir / fname
                    try:
                        _sh.copy2(src, dst)
                        files.append(fname)
                    except Exception:
                        pass
                rep = files[0] if files else None
                if rep:
                    rep_src = out_dir / rep
                    rep_dst = out_dir / "representative.jpg"
                    try:
                        _sh.copy2(rep_src, rep_dst)
                    except Exception:
                        pass
                characters.append({
                    "id": char_id,
                    "name": f"Personatge {ci+1}",
                    "folder": str(out_dir),
                    "num_faces": len(files),
                    "image_url": f"/files/{video_name}/{char_id}/representative.jpg" if rep else "",
                })

            # Escribir analysis.json compatible con 'originales'
            analysis = {
                "caras": [{"embeddings": e} for e in embeddings],
                "voices": [],
                "escenas": [],
            }
            analysis_path = str(base / "analysis.json")
            with open(analysis_path, "w", encoding="utf-8") as f:
                json.dump(analysis, f, ensure_ascii=False)

            face_labels = labels
            num_face_embeddings = len(embeddings)

            print(f"[{job_id}] Personajes detectados: {len(characters)}")
            for char in characters:
                print(f"[{job_id}]   - {char['name']}: {char['num_faces']} caras")
            
            # Enriquecer info de personajes con listado real de imágenes disponibles
            try:
                import glob, os
                for ch in characters:
                    folder = ch.get("folder")
                    face_files = []
                    if folder and os.path.isdir(folder):
                        # soportar patrones face_* y extensiones jpg/png
                        patterns = ["face_*.jpg", "face_*.png"]
                        files = []
                        for pat in patterns:
                            files.extend(glob.glob(os.path.join(folder, pat)))
                        # si no hay face_*, tomar cualquier jpg/png para no dejar vacío
                        if not files:
                            files.extend(glob.glob(os.path.join(folder, "*.jpg")))
                            files.extend(glob.glob(os.path.join(folder, "*.png")))
                        # normalizar nombres de fichero relativos
                        face_files = sorted({os.path.basename(p) for p in files})
                        # Garantizar que representative.(jpg|png) esté el primero si existe
                        for rep_name in ("representative.jpg", "representative.png"):
                            rep_path = os.path.join(folder, rep_name)
                            if os.path.exists(rep_path):
                                if rep_name in face_files:
                                    face_files.remove(rep_name)
                                face_files.insert(0, rep_name)
                    ch["face_files"] = face_files
                    # Ajustar num_faces si hay discrepancia
                    if face_files:
                        ch["num_faces"] = len(face_files)
            except Exception as _e:
                print(f"[{job_id}] WARN - No se pudo enumerar face_files: {_e}")

            # Procesamiento de audio: diarización, ASR y embeddings de voz
            try:
                cfg = load_yaml("config.yaml")
                audio_segments, srt_unmod, full_txt, diar_info, connection_logs = process_audio_for_video(video_path, base, cfg, voice_collection=None)
                # Loggear en consola del engine los eventos de conexión
                try:
                    for ev in (connection_logs or []):
                        msg = ev.get("message") if isinstance(ev, dict) else None
                        if msg:
                            print(f"[{job_id}] {msg}")
                except Exception:
                    pass
            except Exception as e_audio:
                import traceback
                print(f"[{job_id}] WARN - Audio pipeline failed: {e_audio}\n{traceback.format_exc()}")
                audio_segments, srt_unmod, full_txt = [], None, ""
                diar_info = {"diarization_ok": False, "error": str(e_audio)}
                connection_logs = []

            # Fallback: si no hay segmentos de audio, crear uno mínimo del audio completo
            if not audio_segments:
                try:
                    from pathlib import Path as _P
                    from pydub import AudioSegment as _AS
                    wav_out = extract_audio_ffmpeg(video_path, base / f"{_P(video_path).stem}.wav", sr=16000)
                    audio = _AS.from_wav(wav_out)
                    clips_dir = base / "clips"
                    clips_dir.mkdir(parents=True, exist_ok=True)
                    cp = clips_dir / "segment_000.wav"
                    audio.export(cp, format="wav")
                    emb_list = embed_voice_segments([str(cp)])
                    audio_segments = [{
                        "segment": 0,
                        "start": 0.0,
                        "end": float(len(audio) / 1000.0),
                        "speaker": "SPEAKER_00",
                        "text": "",
                        "voice_embedding": emb_list[0] if emb_list else [],
                        "clip_path": str(cp),
                        "lang": "ca",
                        "lang_prob": 1.0,
                    }]
                except Exception as _efb:
                    print(f"[{job_id}] WARN - Audio minimal fallback failed: {_efb}")

            # Clustering jerárquico de voces sobre embeddings válidos
            import numpy as np
            voice_embeddings = [seg.get("voice_embedding") for seg in audio_segments if seg.get("voice_embedding")]
            if voice_embeddings:
                try:
                    Xv = np.array(voice_embeddings)
                    v_labels = hierarchical_cluster_with_min_size(Xv, v_max_groups, v_min_cluster).tolist()
                    print(f"[{job_id}] Clustering jerárquico de voz: {len(set([l for l in v_labels if l >= 0]))} clusters")
                except Exception as _e:
                    print(f"[{job_id}] WARN - Voice clustering failed: {_e}")
                    v_labels = []
            else:
                v_labels = []

            # Guardar resultados primero y luego marcar como completado (evita carreras)
            job["results"] = {
                "characters": characters,
                "num_characters": len(characters),
                "analysis_path": analysis_path,
                "base_dir": str(base),
                "face_labels": face_labels,
                "num_face_embeddings": num_face_embeddings,
                "audio_segments": audio_segments,
                "srt_unmodified": srt_unmod,
                "full_transcription": full_txt,
                "voice_labels": v_labels,
                "num_voice_embeddings": len(voice_embeddings),
                "diarization_info": diar_info,
            }
            job["status"] = JobStatus.DONE
            
            print(f"[{job_id}] DEBUG - job['results'] guardado: {job['results']}")
            
        except Exception as e_detect:
            # Si falla la detección, intentar modo fallback
            import traceback
            print(f"[{job_id}] ✗ Error en detección: {e_detect}")
            print(f"[{job_id}] Traceback: {traceback.format_exc()}")
            print(f"[{job_id}] Usando modo fallback (carpetas vacías)")
            
            # Crear carpetas básicas como fallback
            for sub in ("sources", "faces", "voices", "backgrounds"):
                (base / sub).mkdir(parents=True, exist_ok=True)
            
            # Guardar resultados de fallback y luego marcar como completado
            job["results"] = {
                "characters": [],
                "num_characters": 0,
                "temp_dirs": {
                    "sources": str(base / "sources"),
                    "faces": str(base / "faces"),
                    "voices": str(base / "voices"),
                    "backgrounds": str(base / "backgrounds"),
                },
                "warning": f"Detección falló, usando modo fallback: {str(e_detect)}"
            }
            job["status"] = JobStatus.DONE
        
        print(f"[{job_id}] ✓ Job completado exitosamente")
        
    except Exception as e:
        import traceback
        print(f"[{job_id}] ✗ Error inesperado: {e}")
        try:
            job = jobs.get(job_id)
            if job is not None:
                job["status"] = JobStatus.FAILED
                job["error"] = str(e)
        except Exception:
            pass
        print(f"[{job_id}] Traceback: {traceback.format_exc()}")

@app.post("/generate_audiodescription")
async def generate_audiodescription(video: UploadFile = File(...)):
    try:
        import uuid
        job_id = str(uuid.uuid4())
        vid_name = video.filename or f"video_{job_id}.mp4"
        base = TEMP_ROOT / Path(vid_name).stem

        base.mkdir(parents=True, exist_ok=True)
        # Save temp mp4
        video_path = base / vid_name
        with open(video_path, "wb") as f:
            f.write(await video.read())

        # Run MVP pipeline
        result = ad_generate(str(video_path), base)

        return {
            "status": "done",
            "results": {
                "une_srt": result.get("une_srt", ""),
                "free_text": result.get("free_text", ""),
                "artifacts": result.get("artifacts", {}),
            },
        }
    except Exception as e:
        import traceback
        print(f"/generate_audiodescription error: {e}\n{traceback.format_exc()}")
        raise HTTPException(status_code=500, detail=str(e))

@app.post("/load_casting")
async def load_casting(
    faces_dir: str = Form("identities/faces"),
    voices_dir: str = Form("identities/voices"),
    db_dir: str = Form("chroma_db"),
    drop_collections: bool = Form(False),
):
    client = ensure_chroma(Path(db_dir))
    n_faces = build_faces_index(Path(faces_dir), client, collection_name="index_faces", drop=drop_collections)
    n_voices = build_voices_index(Path(voices_dir), client, collection_name="index_voices", drop=drop_collections)
    return {"ok": True, "faces": n_faces, "voices": n_voices}

@app.post("/finalize_casting")
async def finalize_casting(
    payload: dict = Body(...),
):
    """
    Consolidate selected face and voice clusters into identities directories and build indices.
    Expected payload:
    {
      "video_name": str,
      "base_dir": str,  # engine temp base for this video
      "characters": [
        {"id": "char1", "name": "Nom", "folder": "/tmp/temp/<video>/char1", "kept_files": ["representative.jpg", ...], "description": "..."}, ...
      ],
      "voice_clusters": [
        {"label": 0, "name": "SPEAKER_00", "clips": ["segment_000.wav", ...]}, ...
      ]
    }
    """
    import os
    import shutil
    from pathlib import Path as _P

    video_name = payload.get("video_name")
    base_dir = payload.get("base_dir")
    characters = payload.get("characters", []) or []
    voice_clusters = payload.get("voice_clusters", []) or []

    if not video_name or not base_dir:
        raise HTTPException(status_code=400, detail="Missing video_name or base_dir")

    faces_out = IDENTITIES_ROOT / video_name / "faces"
    voices_out = IDENTITIES_ROOT / video_name / "voices"
    faces_out.mkdir(parents=True, exist_ok=True)
    voices_out.mkdir(parents=True, exist_ok=True)

    # Consolidate faces per character name (merge same names)
    for ch in characters:
        ch_name = (ch.get("name") or "Unknown").strip() or "Unknown"
        ch_folder = ch.get("folder")
        kept = ch.get("kept_files") or []
        if not ch_folder or not os.path.isdir(ch_folder):
            continue
        dst_dir = faces_out / ch_name
        dst_dir.mkdir(parents=True, exist_ok=True)
        for fname in kept:
            src = _P(ch_folder) / fname
            if src.exists() and src.is_file():
                try:
                    shutil.copy2(src, dst_dir / fname)
                except Exception:
                    pass

    # Consolidate voices per cluster name
    clips_dir = _P(base_dir) / "clips"
    for vc in voice_clusters:
        v_name = (vc.get("name") or f"SPEAKER_{int(vc.get('label',0)):02d}").strip()
        dst_dir = voices_out / v_name
        dst_dir.mkdir(parents=True, exist_ok=True)
        for wav in (vc.get("clips") or []):
            src = clips_dir / wav
            if src.exists() and src.is_file():
                try:
                    shutil.copy2(src, dst_dir / wav)
                except Exception:
                    pass

    # Build indices using casting_loader helpers
    db_dir = IDENTITIES_ROOT / video_name / "chroma_db"
    client = ensure_chroma(db_dir)
    n_faces = build_faces_index(faces_out, client, collection_name="index_faces", deepface_model='Facenet512', drop=True)
    n_voices = build_voices_index(voices_out, client, collection_name="index_voices", drop=True)

    # Summary of identities
    face_identities = sorted([p.name for p in faces_out.iterdir() if p.is_dir()]) if faces_out.exists() else []
    voice_identities = sorted([p.name for p in voices_out.iterdir() if p.is_dir()]) if voices_out.exists() else []

    return {
        "ok": True,
        "video_name": video_name,
        "faces_dir": str(faces_out),
        "voices_dir": str(voices_out),
        "db_dir": str(db_dir),
        "n_faces_embeddings": n_faces,
        "n_voices_embeddings": n_voices,
        "face_identities": face_identities,
        "voice_identities": voice_identities,
    }

@app.get("/files_scene/{video_name}/{scene_id}/{filename}")
def serve_scene_file(video_name: str, scene_id: str, filename: str):
    file_path = TEMP_ROOT / video_name / "scenes" / scene_id / filename
    if not file_path.exists():
        raise HTTPException(status_code=404, detail="File not found")
    return FileResponse(file_path)

@app.post("/detect_scenes")
async def detect_scenes(
    video: UploadFile = File(...),
    max_groups: int = Form(default=3),
    min_cluster_size: int = Form(default=3),
    frame_interval_sec: float = Form(default=0.5),
):
    """
    Detecta clústers d'escenes mitjançant clustering jeràrquic d'histogrames de color.
    Retorna una llista de scene_clusters estructurada de forma similar a characters.
    """
    import cv2
    import numpy as np

    # Guardar el vídeo temporalment
    video_name = Path(video.filename).stem
    dst_video = VIDEOS_ROOT / f"{video_name}.mp4"
    with dst_video.open("wb") as f:
        shutil.copyfileobj(video.file, f)

    cap = cv2.VideoCapture(str(dst_video))
    if not cap.isOpened():
        raise HTTPException(status_code=400, detail="Cannot open video")

    fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
    step = max(1, int(frame_interval_sec * fps))

    frames = []
    metas = []
    idx = 0
    while True:
        ret = cap.grab()
        if not ret:
            break
        if idx % step == 0:
            ret2, frame = cap.retrieve()
            if not ret2:
                break
            # Reduir mida per estabilitat i càlcul ràpid
            small = cv2.resize(frame, (160, 90))
            hsv = cv2.cvtColor(small, cv2.COLOR_BGR2HSV)
            # Histograma per canal
            h_hist = cv2.calcHist([hsv],[0],None,[32],[0,180]).flatten()
            s_hist = cv2.calcHist([hsv],[1],None,[32],[0,256]).flatten()
            v_hist = cv2.calcHist([hsv],[2],None,[32],[0,256]).flatten()
            hist = np.concatenate([h_hist, s_hist, v_hist])
            hist = hist / (np.linalg.norm(hist) + 1e-8)
            frames.append(hist)
            metas.append({"index": idx, "time_sec": idx/float(fps)})
        idx += 1
    cap.release()

    if not frames:
        return {"scene_clusters": []}

    X = np.array(frames)
    labels = hierarchical_cluster_with_min_size(X, max_groups, min_cluster_size).tolist()
    print(f"Scene clustering jeràrquic: {len(set([l for l in labels if l >= 0]))} clusters")

    # Agrupar per etiqueta (>=0)
    clusters = {}
    for i, lbl in enumerate(labels):
        if lbl is None or lbl < 0:
            continue
        clusters.setdefault(int(lbl), []).append(i)

    # Escriure imatges representatives per a cada clúster
    base = TEMP_ROOT / video_name / "scenes"
    base.mkdir(parents=True, exist_ok=True)
    scene_list = []
    cap = cv2.VideoCapture(str(dst_video))
    for lbl, idxs in sorted(clusters.items(), key=lambda x: x[0]):
        scene_id = f"scene_{int(lbl):02d}"
        out_dir = base / scene_id
        out_dir.mkdir(parents=True, exist_ok=True)
        frame_files = []
        # Guardar fins a 12 frames per clúster
        for k, fi in enumerate(idxs[:12]):
            frame_num = metas[fi]["index"]
            cap.set(cv2.CAP_PROP_POS_FRAMES, frame_num)
            ret2, frame = cap.read()
            if not ret2:
                continue
            fn = f"frame_{k:03d}.jpg"
            cv2.imwrite(str(out_dir / fn), frame)
            frame_files.append(fn)
        # Representative
        rep = frame_files[0] if frame_files else None
        image_url = f"/files_scene/{video_name}/{scene_id}/{rep}" if rep else ""
        
        # Llamar a svision para describir la escena representativa
        scene_description = ""
        scene_name = f"Escena {lbl+1}"
        if rep:
            rep_full_path = out_dir / rep
            if rep_full_path.exists():
                print(f"Llamando a svision para describir {scene_id}...")
                try:
                    scene_description, scene_name = describe_image_with_svision(str(rep_full_path), is_face=False)
                    if not scene_name:
                        scene_name = f"Escena {lbl+1}"
                except Exception as e:
                    print(f"Error describiendo {scene_id}: {e}")
        
        scene_list.append({
            "id": scene_id,
            "name": scene_name,
            "description": scene_description,
            "folder": str(out_dir),
            "num_frames": len(frame_files),
            "image_url": image_url,
            "frame_files": frame_files,
        })
    cap.release()

    return {"scene_clusters": scene_list, "base_dir": str(base)}

@app.post("/refine_narration")
async def refine_narration(
    dialogues_srt: str = Form(...),
    frame_descriptions_json: str = Form("[]"),
    config_path: str = Form("config.yaml"),
):
    cfg = load_yaml(config_path)
    frames = json.loads(frame_descriptions_json)
    model_name = cfg.get("narration", {}).get("model", "salamandra-instruct")
    use_remote = model_name in (cfg.get("models", {}).get("routing", {}).get("use_remote_for", []))

    if use_remote:
        router = LLMRouter(cfg)
        system_msg = (
            "Eres un sistema de audiodescripción que cumple UNE-153010. "
            "Fusiona diálogos del SRT con descripciones concisas en los huecos, evitando redundancias. "
            "Devuelve JSON con {narrative_text, srt_text}."
        )
        prompt = json.dumps({"dialogues_srt": dialogues_srt, "frames": frames, "rules": cfg.get("narration", {})}, ensure_ascii=False)
        try:
            txt = router.instruct(prompt=prompt, system=system_msg, model=model_name)
            out = {}
            try:
                out = json.loads(txt)
            except Exception:
                out = {"narrative_text": txt, "srt_text": ""}
            return {
                "narrative_text": out.get("narrative_text", ""),
                "srt_text": out.get("srt_text", ""),
                "approved": True,
                "critic_feedback": "",
            }
        except Exception:
            ns = NarrationSystem(model_url=None, une_guidelines_path=cfg.get("narration", {}).get("narration_une_guidelines_path", "UNE_153010.txt"))
            res = ns.run(dialogues_srt, frames)
            return {"narrative_text": res.narrative_text, "srt_text": res.srt_text, "approved": res.approved, "critic_feedback": res.critic_feedback}

    ns = NarrationSystem(model_url=None, une_guidelines_path=cfg.get("narration", {}).get("une_guidelines_path", "UNE_153010.txt"))
    out = ns.run(dialogues_srt, frames)
    return {"narrative_text": out.narrative_text, "srt_text": out.srt_text, "approved": out.approved, "critic_feedback": out.critic_feedback}

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860)