app/services/embedder_adapter.py

"""Embedder adapter using sentence-transformers — fully non-blocking."""
from app.ports.embedder import EmbedderPort
from sentence_transformers import SentenceTransformer
from typing import List
from app.config import get_settings
import asyncio
import logging

logger = logging.getLogger(__name__)
settings = get_settings()

# Shared thread-pool executor for CPU-bound work
_executor = None

def _get_executor():
    """Lazy-init a dedicated thread pool for embedding work."""
    global _executor
    if _executor is None:
        from concurrent.futures import ThreadPoolExecutor
        # Keep 1 worker — the model isn't thread-safe for parallel calls
        _executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="embedder")
    return _executor


class SentenceTransformerAdapter(EmbedderPort):
    """Sentence-transformers implementation — runs in executor to avoid blocking."""

    def __init__(self):
        logger.info(f"Loading embedding model: {settings.EMBEDDING_MODEL}")
        self.model = SentenceTransformer(settings.EMBEDDING_MODEL)
        self.dimension = settings.EMBEDDING_DIMENSION

    async def embed_text(self, text: str) -> List[float]:
        """Embed a single text without blocking the event loop."""
        loop = asyncio.get_event_loop()
        result = await loop.run_in_executor(
            _get_executor(),
            lambda: self.model.encode(text, convert_to_numpy=True).tolist()
        )
        return result

    async def embed_batch(self, texts: List[str]) -> List[List[float]]:
        """Embed a batch of texts without blocking the event loop."""
        loop = asyncio.get_event_loop()
        result = await loop.run_in_executor(
            _get_executor(),
            lambda: self.model.encode(
                texts,
                convert_to_numpy=True,
                show_progress_bar=True,
                batch_size=8,          # smaller batches = more responsive
            ).tolist()
        )
        return result

    def get_dimension(self) -> int:
        return self.dimension