blizzarman/polyglot-tutor

1"""Centralised, environment-driven configuration.2 3Every provider choice (LLM, ASR, TTS, storage) is a plain env var so the same4image runs as the free HF Space ("light" mode) or against the local GPU box5("premium" mode) without code changes.6 7The `Literal` types are intentionally restricted to *implemented* providers:8adding an implementation means widening the Literal, so the config can never9silently point at a backend that does not exist yet.10"""11 12from functools import lru_cache13from typing import Literal14 15from pydantic import SecretStr16from pydantic_settings import BaseSettings, SettingsConfigDict17 18 19class Settings(BaseSettings):20    model_config = SettingsConfigDict(21        env_file=".env",22        env_file_encoding="utf-8",23        extra="ignore",24    )25 26    # --- App ---27    app_env: Literal["dev", "prod"] = "dev"28    log_level: str = "INFO"29 30    # --- Languages ---31    default_source_lang: str = "fr"32    default_target_lang: str = "en"33 34    # --- LLM ---35    llm_provider: Literal["fake", "gemini", "openai", "mistral", "ollama"] = "fake"36    llm_model: str = "gemini-2.5-flash"37    llm_api_key: SecretStr | None = None38    llm_base_url: str | None = None39    llm_timeout_s: float = 30.040 41    # --- CEFR classifier (M1) ---42    cefr_model_path: str | None = None  # local ONNX artifact dir (dev) — takes precedence43    cefr_model_id: str | None = None  # HF model repo id (Space; HF_TOKEN env for private repos)44    cache_dir: str = ".cache/tutor"  # content-addressed cache for LLM products45 46    # --- ASR (M2) / TTS / storage ---47    asr_provider: Literal["fake", "faster_whisper"] = "fake"48    asr_model: str = "small"  # faster-whisper size; see docs/evals/m2_asr_latency.md49    asr_compute_type: str = "int8"50    asr_cpu_threads: int = 251    tts_provider: Literal["fake"] = "fake"  # M2 TTS is browser-side (Web Speech API)52    storage_backend: Literal["memory"] = "memory"53 54    # --- Gradio ---55    host: str = "0.0.0.0"56    port: int = 786057    gradio_auth_username: str | None = None58    gradio_auth_password: SecretStr | None = None59 60 61@lru_cache62def get_settings() -> Settings:63    """Process-wide settings singleton (tests build `Settings` directly instead)."""64    return Settings()65