services/recommend.py

import asyncio
import unicodedata
import re
import numpy as np
from datetime import datetime
from typing import List, Dict, Optional, Set, Tuple

from services.tmdb import get_genre_map, fetch_by_genres, fetch_popular, search_movie, TMDB_IMG
from services.embedding import encode_movie, encode_mood
from services.engine import EngineRecommandationGroupe

MIN_POPULARITY = 8.0

MOOD_KEYWORDS: Dict[str, List[str]] = {
    "action": ["Action", "Thriller"],
    "comedie": ["Comedy"], "comédie": ["Comedy"],
    "drole": ["Comedy"], "drôle": ["Comedy"], "rire": ["Comedy"],
    "legere": ["Comedy"], "légère": ["Comedy"],
    "horreur": ["Horror"], "peur": ["Horror"],
    "drame": ["Drama"], "drama": ["Drama"],
    "emouvant": ["Drama", "Romance"], "poignant": ["Drama"], "intense": ["Thriller", "Drama"],
    "triste": ["Drama"], "melancolique": ["Drama"], "mélancolique": ["Drama"],
    "romantique": ["Romance"], "romance": ["Romance"], "amour": ["Romance"],
    "aventure": ["Adventure"],
    "science-fiction": ["Science Fiction"], "sf": ["Science Fiction"],
    "fantaisie": ["Fantasy"], "fantasy": ["Fantasy"],
    "animation": ["Animation"],
    "famille": ["Family"], "enfants": ["Family", "Animation"],
    "documentaire": ["Documentary"],
    "policier": ["Crime", "Mystery"], "crime": ["Crime"],
    "mystere": ["Mystery"], "mystère": ["Mystery"],
    "guerre": ["War"], "histoire": ["History"],
    "musical": ["Music"], "western": ["Western"],
    "thriller": ["Thriller"], "suspense": ["Thriller", "Mystery"],
    "feel-good": ["Comedy", "Romance"], "feel good": ["Comedy", "Romance"],
    "popcorn": ["Action", "Adventure"], "detente": ["Comedy", "Family"],
    "frissons": ["Horror", "Thriller"],
}


def _normalize(key: str) -> str:
    key = key.lower()
    key = unicodedata.normalize("NFD", key)
    key = "".join(c for c in key if unicodedata.category(c) != "Mn")
    key = re.sub(r"[^\w\s_]", "", key)
    return re.sub(r"\s+", " ", key).strip()


def extract_genres_from_mood(mood_text: str) -> List[str]:
    mood_norm = _normalize(mood_text)
    genres: set = set()
    for keyword, genre_list in MOOD_KEYWORDS.items():
        if _normalize(keyword) in mood_norm:
            genres.update(genre_list)
    return list(genres)


async def _fetch_and_embed(film: Dict, genre_map: Dict) -> Optional[tuple]:
    try:
        result = await search_movie(film["name"], film["year"])
        if result and result.get("overview"):
            genres = [genre_map.get(gid, "") for gid in result.get("genre_ids", [])]
            genres = [g for g in genres if g]
            emb = encode_movie(result["overview"], genres)
            return film["key"], emb
    except Exception:
        pass
    return None


async def build_profile_embeddings(
    history: List[Dict], genre_map: Dict, top_k: int = 25
) -> Dict[str, np.ndarray]:
    top_films = sorted(history, key=lambda x: x["rating"], reverse=True)[:top_k]
    results = await asyncio.gather(
        *[_fetch_and_embed(f, genre_map) for f in top_films],
        return_exceptions=True,
    )
    out = {}
    for r in results:
        if r and not isinstance(r, Exception):
            key, emb = r
            out[key] = emb
    return out


async def get_watchlist_films_for_mood(
    watchlist_sets: List[Set[str]],
    genres_mood: List[str],
    genre_map: Dict[int, str],
    max_guaranteed: int = 2,
) -> Tuple[List[Dict], List[Dict]]:
    """
    Retourne (films_garantis, films_boostes).
    - Garantis : TOUS les utilisateurs ont le film ET il correspond au mood.
    - Boostes  : CERTAINS utilisateurs ont le film ET il correspond au mood.
    """
    all_keys: Set[str] = set()
    for ws in watchlist_sets:
        all_keys.update(ws)

    if not all_keys:
        return [], []
    n_users = len(watchlist_sets)

    async def check(raw_key: str) -> Optional[Dict]:
        try:
            parts = raw_key.rsplit("_", 1)
            name = parts[0]
            year = int(parts[1]) if len(parts) > 1 and parts[1].isdigit() else None
            result = await search_movie(name, year)
            if not result or not result.get("overview"):
                return None
            title = result["title"].strip()
            raw_year = result.get("release_date", "")[:4]
            r_year = int(raw_year) if raw_year.isdigit() else None
            key = (title + "_" + str(r_year)) if r_year else title
            genres = [genre_map.get(gid, "") for gid in result.get("genre_ids", [])]
            genres = [g for g in genres if g]
            # Seuls les films dont un genre correspond au mood sont eligibles
            if genres_mood and not any(g in genres_mood for g in genres):
                return None
            return {
                "key": key,
                "tmdb_id": int(result["id"]),
                "title": title,
                "year": r_year,
                "overview": result.get("overview", ""),
                "genres": genres,
                "popularity": float(result.get("popularity", 0.0)),
                "poster_path": TMDB_IMG + result["poster_path"] if result.get("poster_path") else None,
                "on_watchlist": True,
                "score": 0.97,
                "_user_count": sum(1 for ws in watchlist_sets if raw_key in ws),
                "_all_users": sum(1 for ws in watchlist_sets if raw_key in ws) >= n_users,
            }
        except Exception:
            return None

    results = await asyncio.gather(
        *[check(k) for k in list(all_keys)[:50]],
        return_exceptions=True,
    )
    matches = [r for r in results if r and not isinstance(r, Exception)]
    guaranteed = sorted([m for m in matches if m["_all_users"]], key=lambda x: x["popularity"], reverse=True)
    partial = sorted([m for m in matches if not m["_all_users"]], key=lambda x: x["popularity"], reverse=True)
    return guaranteed[:max_guaranteed], partial


async def run_pipeline(
    histories: List[List[Dict]],
    mood_text: str,
    top_n: int = 5,
    watchlist_sets: Optional[List[Set[str]]] = None,
) -> List[Dict]:
    seen_norm = {
        _normalize(film["key"])
        for history in histories
        for film in history
    }

    genre_map = await get_genre_map()
    genres_mood = extract_genres_from_mood(mood_text)

    # 1. Pool de candidats : genre-specifique ou populaire en fallback
    genre_pool, popular_pool = await asyncio.gather(
        fetch_by_genres(seen_norm, genre_map, genres_mood, max_pages=10),
        fetch_popular(seen_norm, genre_map, max_pages=3),
    )
    all_candidates = genre_pool if genre_pool else popular_pool

    # Filtre popularite : retire les films trop obscurs
    candidates = [c for c in all_candidates.values() if c.get("popularity", 0) >= MIN_POPULARITY]
    if not candidates:
        return []

    # 2. Embeddings des candidats
    candidate_embeddings: Dict[str, np.ndarray] = {
        c["key"]: encode_movie(c["overview"], c["genres"]) for c in candidates
    }

    # 3. Profils utilisateurs
    profile_embs_list = await asyncio.gather(
        *[build_profile_embeddings(h, genre_map) for h in histories]
    )

    engine = EngineRecommandationGroupe(
        omega_hist=0.35,
        omega_mood=0.65,
        omega_pen=0.2,
        delta=0.35,
        gamma_pop=0.55,
    )

    profils_groupe = []
    for history, emb_catalogue in zip(histories, profile_embs_list):
        if emb_catalogue:
            profil = engine.calcul_profil_utilisateur(history, emb_catalogue)
            profils_groupe.append(profil)

    if not profils_groupe:
        return []

    v_mood = encode_mood(mood_text)

    # 4. Scoring
    scored = []
    for c in candidates:
        if c["key"] not in candidate_embeddings:
            continue
        score = engine.scoring_film_candidat(
            v_m=candidate_embeddings[c["key"]],
            genres_m=c["genres"],
            pop_m=c["popularity"],
            profils_groupe=profils_groupe,
            v_mood=v_mood,
            genres_mood=genres_mood,
        )
        scored.append({**c, "score": round(score, 4), "on_watchlist": False})

    scored.sort(key=lambda x: x["score"], reverse=True)
    results = scored[:top_n]

    # 5. Injection watchlist
    if watchlist_sets:
        wl_guaranteed, wl_partial = await get_watchlist_films_for_mood(
            watchlist_sets, genres_mood, genre_map
        )
        if wl_guaranteed:
            wl_ids = {f["tmdb_id"] for f in wl_guaranteed}
            results = [r for r in results if r["tmdb_id"] not in wl_ids]
            results = (wl_guaranteed + results)[:top_n]
        if wl_partial:
            partial_ids = {f["tmdb_id"] for f in wl_partial}
            changed = False
            for r in results:
                if r["tmdb_id"] in partial_ids:
                    r["score"] = round(r["score"] * 1.3, 4)
                    r["on_watchlist"] = True
                    changed = True
            if changed:
                results = sorted(results, key=lambda x: x["score"], reverse=True)[:top_n]

    return results


async def run_pipeline_from_text(
    liked_film_names: List[str],
    mood_text: str,
    top_n: int = 5,
) -> Tuple[List[Dict], List[str]]:
    """
    Pipeline pour la saisie manuelle.
    Meme algorithme que Letterboxd mais avec un historique synthetique.
    """
    genre_map = await get_genre_map()

    search_results = await asyncio.gather(
        *[search_movie(name, None) for name in liked_film_names],
        return_exceptions=True,
    )

    synthetic_history = []
    profile_embeddings: Dict[str, np.ndarray] = {}
    films_found: List[str] = []

    for result in search_results:
        if not result or isinstance(result, Exception) or not result.get("title"):
            continue
        title = result["title"].strip()
        raw_year = result.get("release_date", "")[:4]
        year = int(raw_year) if raw_year.isdigit() else None
        key = (title + "_" + str(year)) if year else title
        genres = [genre_map.get(gid, "") for gid in result.get("genre_ids", [])]
        genres = [g for g in genres if g]
        films_found.append(title)
        if result.get("overview"):
            profile_embeddings[key] = encode_movie(result["overview"], genres)
        synthetic_history.append({
            "key": key, "name": title, "year": year,
            "rating": 4.5, "date_viewed": datetime.now(),
        })

    if not synthetic_history or not profile_embeddings:
        raise ValueError("Aucun film trouve sur TMDB. Verifie les titres.")

    genres_mood = extract_genres_from_mood(mood_text)
    seen_norm: Set[str] = set()

    genre_pool, popular_pool = await asyncio.gather(
        fetch_by_genres(seen_norm, genre_map, genres_mood, max_pages=10),
        fetch_popular(seen_norm, genre_map, max_pages=3),
    )
    all_candidates = genre_pool if genre_pool else popular_pool
    candidates = [c for c in all_candidates.values() if c.get("popularity", 0) >= MIN_POPULARITY]

    if not candidates:
        return [], films_found

    candidate_embeddings: Dict[str, np.ndarray] = {
        c["key"]: encode_movie(c["overview"], c["genres"]) for c in candidates
    }

    # Poids plus equilibres pour la saisie manuelle (peu de films = profil moins riche)
    engine = EngineRecommandationGroupe(
        omega_hist=0.4,
        omega_mood=0.6,
        omega_pen=0.1,
        delta=0.35,
        gamma_pop=0.55,
    )

    profil = engine.calcul_profil_utilisateur(synthetic_history, profile_embeddings)
    v_mood = encode_mood(mood_text)

    scored = []
    for c in candidates:
        if c["key"] not in candidate_embeddings:
            continue
        score = engine.scoring_film_candidat(
            v_m=candidate_embeddings[c["key"]],
            genres_m=c["genres"],
            pop_m=c["popularity"],
            profils_groupe=[profil],
            v_mood=v_mood,
            genres_mood=genres_mood,
        )
        scored.append({**c, "score": round(score, 4), "on_watchlist": False})

    scored.sort(key=lambda x: x["score"], reverse=True)
    return scored[:top_n], films_found