from __future__ import annotations

from typing import List, Optional, TYPE_CHECKING

from app.infrastructure.settings import EMBEDDINGS_MODEL

if TYPE_CHECKING:  # pragma: no cover - typing only, no runtime cost
    from sentence_transformers import SentenceTransformer

_model_singleton: Optional["SentenceTransformer"] = None


def get_device() -> str:
    """Prefer CUDA if available, otherwise CPU.

    Imports torch lazily: on an 8GB RAM / CPU-only laptop there is no
    benefit to importing torch (and paying its startup/memory cost) until
    an embedding is actually requested, so the FastAPI process can boot
    and answer /api/health almost instantly.
    """
    import torch  # local import - see docstring

    return "cuda" if torch.cuda.is_available() else "cpu"


def get_embedding_model() -> "SentenceTransformer":
    global _model_singleton
    if _model_singleton is None:
        from sentence_transformers import SentenceTransformer  # local import - see get_device()

        device = get_device()
        _model_singleton = SentenceTransformer(EMBEDDINGS_MODEL, device=device)
    return _model_singleton


def embed_texts(texts: List[str]) -> List[List[float]]:
    """Embed a batch of texts into normalized 384-dim vectors (MiniLM-L6-v2).

    Normalized so that pgvector's cosine-distance operator (`<=>`) behaves
    consistently for the RAG retrieval step in `app.services.vector_store`.
    """
    if not texts:
        return []
    model = get_embedding_model()
    embeddings = model.encode(
        texts,
        batch_size=32,
        normalize_embeddings=True,
        convert_to_numpy=True,
        show_progress_bar=False,
    )
    return embeddings.tolist()