from __future__ import annotations
from typing import List
from functools import lru_cache

from .config import get_settings

# Lazy import so startup stays fast
@lru_cache
def _load_model():
    from sentence_transformers import SentenceTransformer
    settings = get_settings()
    model = SentenceTransformer(settings.MODEL_NAME, device="cpu")
    return model

def embed_texts(texts: List[str]) -> list[list[float]]:
    model = _load_model()
    # Normalize to list of str
    texts = [t if isinstance(t, str) else str(t) for t in texts]
    vecs = model.encode(texts, normalize_embeddings=True, convert_to_numpy=False)
    # Ensure pure-python list of floats
    return [list(map(float, v)) for v in vecs]