"""
Self-hosted text embeddings (fastembed - ONNX runtime, CPU, no API key, no
external network call at runtime) for document RAG: turning uploaded text
into vectors stored in MongoDB Atlas, and turning the user's question into a
vector to search against them.

Same reasoning as Piper for voice: a self-hosted model has no external
service to be rate-limited or blocked by, which matters on a cloud host.
"""
import asyncio
from functools import lru_cache

EMBEDDING_MODEL_NAME = "BAAI/bge-small-en-v1.5"  # 384 dimensions, small, fast on CPU


@lru_cache(maxsize=1)
def _get_model():
    from fastembed import TextEmbedding

    return TextEmbedding(model_name=EMBEDDING_MODEL_NAME)


def _embed_sync(texts: list[str]) -> list[list[float]]:
    model = _get_model()
    return [vec.tolist() for vec in model.embed(texts)]


async def embed_texts(texts: list[str]) -> list[list[float]]:
    return await asyncio.get_event_loop().run_in_executor(None, _embed_sync, texts)


async def embed_text(text: str) -> list[float]:
    vectors = await embed_texts([text])
    return vectors[0]