"""Centralised, environment-driven configuration.

Every provider choice (LLM, ASR, TTS, storage) is a plain env var so the same
image runs as the free HF Space ("light" mode) or against the local GPU box
("premium" mode) without code changes.

The `Literal` types are intentionally restricted to *implemented* providers:
adding an implementation means widening the Literal, so the config can never
silently point at a backend that does not exist yet.
"""

from functools import lru_cache
from typing import Literal

from pydantic import SecretStr
from pydantic_settings import BaseSettings, SettingsConfigDict


class Settings(BaseSettings):
    model_config = SettingsConfigDict(
        env_file=".env",
        env_file_encoding="utf-8",
        extra="ignore",
    )

    # --- App ---
    app_env: Literal["dev", "prod"] = "dev"
    log_level: str = "INFO"

    # --- Languages ---
    default_source_lang: str = "fr"
    default_target_lang: str = "en"

    # --- LLM ---
    llm_provider: Literal["fake", "gemini", "openai", "mistral", "ollama"] = "fake"
    llm_model: str = "gemini-2.5-flash"
    llm_api_key: SecretStr | None = None
    llm_base_url: str | None = None
    llm_timeout_s: float = 30.0

    # --- CEFR classifier (M1) ---
    cefr_model_path: str | None = None  # local ONNX artifact dir (dev) — takes precedence
    cefr_model_id: str | None = None  # HF model repo id (Space; HF_TOKEN env for private repos)
    cache_dir: str = ".cache/tutor"  # content-addressed cache for LLM products

    # --- ASR (M2) / TTS / storage ---
    asr_provider: Literal["fake", "faster_whisper"] = "fake"
    asr_model: str = "small"  # faster-whisper size; see docs/evals/m2_asr_latency.md
    asr_compute_type: str = "int8"
    asr_cpu_threads: int = 2
    tts_provider: Literal["fake"] = "fake"  # M2 TTS is browser-side (Web Speech API)
    storage_backend: Literal["memory"] = "memory"

    # --- Gradio ---
    host: str = "0.0.0.0"
    port: int = 7860
    gradio_auth_username: str | None = None
    gradio_auth_password: SecretStr | None = None


@lru_cache
def get_settings() -> Settings:
    """Process-wide settings singleton (tests build `Settings` directly instead)."""
    return Settings()