app/services/nlp_client.py

from __future__ import annotations

import asyncio
import logging
from typing import Any

import httpx
import numpy as np

from app.config import get_settings

logger = logging.getLogger(__name__)

FAISS_DIM = 384  # match Django's MiniLM output dim (NewsArticle.search_vector)

_NLP_MAX_RETRIES = 3
_NLP_TIMEOUT_S = 60.0  # generous — accounts for HF Space cold start


def _nlp_service_url() -> str:
    settings = get_settings()
    url = settings.nlp_service_url.rstrip("/")
    if not url:
        raise RuntimeError("NLP_SERVICE_URL not configured in settings")
    return url


def _nlp_headers() -> dict[str, str]:
    settings = get_settings()
    if not settings.nlp_service_token:
        raise RuntimeError("NLP_SERVICE_TOKEN not configured in settings")
    return {
        "Authorization": f"Bearer {settings.nlp_service_token}",
        "Content-Type": "application/json",
    }


async def _nlp_post(path: str, payload: dict) -> Any:
    """POST to the shared NLP microservice with retry on timeout/5xx (cold-start safe)."""
    url = f"{_nlp_service_url()}{path}"
    headers = _nlp_headers()
    last_exc: Exception | None = None

    async with httpx.AsyncClient(timeout=_NLP_TIMEOUT_S) as client:
        for attempt in range(1, _NLP_MAX_RETRIES + 1):
            try:
                resp = await client.post(url, json=payload, headers=headers)
                resp.raise_for_status()
                return resp.json()
            except (httpx.TimeoutException, httpx.ConnectError) as exc:
                last_exc = exc
                logger.warning(
                    "NLP service attempt %d/%d timed out: %s",
                    attempt, _NLP_MAX_RETRIES, exc,
                )
                if attempt < _NLP_MAX_RETRIES:
                    await asyncio.sleep(2 ** attempt)  # 2s, 4s
            except httpx.HTTPStatusError as exc:
                if exc.response.status_code < 500:
                    raise  # 4xx — don't retry
                last_exc = exc
                logger.warning(
                    "NLP service attempt %d/%d HTTP error: %s",
                    attempt, _NLP_MAX_RETRIES, exc,
                )
                if attempt < _NLP_MAX_RETRIES:
                    await asyncio.sleep(2 ** attempt)

    raise RuntimeError(
        f"NLP microservice unavailable after {_NLP_MAX_RETRIES} attempts"
    ) from last_exc


async def embed(texts: list[str]) -> np.ndarray:
    """
    Embed texts via the shared remote MiniLM endpoint (same HF Space Django uses).
    Returns float32 array of shape (len(texts), FAISS_DIM), L2-normalised.
    """
    result: list[list[float]] = await _nlp_post("/embed", {"texts": texts})
    return np.array(result, dtype="float32")