app/core/config.py

from functools import lru_cache
from pathlib import Path

from pydantic import AliasChoices, Field
from pydantic_settings import BaseSettings, SettingsConfigDict


BACKEND_DIR = Path(__file__).resolve().parents[2]
PROJECT_ROOT = BACKEND_DIR.parent


class Settings(BaseSettings):
    app_name: str = "AI Exam Success API"
    app_version: str = "0.1.0"
    environment: str = "development"
    database_url: str = f"sqlite:///{BACKEND_DIR / 'exam_success_dev.db'}"
    # Dev defaults stay small; production .env should raise these for 1k+ concurrent students.
    database_pool_size: int = Field(
        default=2,
        validation_alias=AliasChoices("DATABASE_POOL_SIZE"),
    )
    database_max_overflow: int = Field(
        default=3,
        validation_alias=AliasChoices("DATABASE_MAX_OVERFLOW"),
    )
    database_pool_timeout_seconds: int = Field(
        default=10,
        validation_alias=AliasChoices("DATABASE_POOL_TIMEOUT_SECONDS"),
    )
    database_pool_recycle_seconds: int = Field(
        default=1800,
        validation_alias=AliasChoices("DATABASE_POOL_RECYCLE_SECONDS"),
    )
    cors_origins: str = "http://localhost:3000,http://127.0.0.1:3000,http://localhost:3001,http://127.0.0.1:3001,http://localhost:3003,http://127.0.0.1:3003"
    upload_dir: str = "uploads"
    ai_provider: str = "openai"
    # The tuition brain uses its OWN text provider so it can run on a cheap
    # text-adapter (Cloudflare Workers AI) without forcing the rest of the app
    # off its main provider. Must be a text-adapter name
    # (cloudflare_workers_ai | openrouter | nvidia_nim); anything else -> the
    # brain falls back to its deterministic plan.
    tuition_brain_provider: str = Field(
        default="cloudflare_workers_ai",
        validation_alias=AliasChoices("TUITION_BRAIN_PROVIDER"),
    )
    # --- OpenAI (fast + high-quality primary provider) ---
    openai_api_key: str | None = Field(
        default=None,
        validation_alias=AliasChoices("OPENAI_API_KEY"),
    )
    openai_base_url: str = "https://api.openai.com/v1"
    openai_model_main: str = "gpt-5-mini"
    openai_model_heavy: str = "gpt-5.2"
    openai_timeout_seconds: int = 60
    openai_max_retries: int = 1
    openai_reasoning_effort_fast: str = "minimal"
    openai_reasoning_effort_heavy: str = "low"
    # --- Sarvam AI (primary production provider) ---
    sarvam_api_key: str | None = Field(
        default=None,
        validation_alias=AliasChoices("SARVAM_API_KEY"),
    )
    sarvam_base_url: str = "https://api.sarvam.ai/v1"
    sarvam_model_main: str = "sarvam-30b"     # sarvam-m deprecated; 30b is the fast reasoning model
    sarvam_model_heavy: str = "sarvam-30b"    # reasoning model for deep analysis
    sarvam_model_notes: str = "sarvam-105b"        # upgraded: more tokens for richer notes
    sarvam_model_quiz: str = "sarvam-105b"          # upgraded: more questions, better quality
    sarvam_model_flashcards: str = "sarvam-105b"    # upgraded: more cards, richer back-sides
    sarvam_model_exam_mode: str = "sarvam-105b"
    sarvam_model_pyq_analysis: str = "sarvam-105b"
    sarvam_model_video_script: str = "sarvam-105b"
    sarvam_timeout_seconds: int = 90
    sarvam_max_retries: int = 2
    # --- OpenRouter (legacy provider — kept for backward compat) ---
    openrouter_api_key: str | None = Field(
        default=None,
        validation_alias=AliasChoices("OPENROUTER_API_KEY"),
    )
    openrouter_base_url: str = "https://openrouter.ai/api/v1"
    # User-facing fallbacks must support structured outputs and avoid unstable
    # free-tier routing. The old DeepSeek free slug now returns HTTP 404.
    openrouter_model_main: str = "deepseek/deepseek-v4-flash"
    openrouter_model_llama: str = "qwen/qwen3-next-80b-a3b-instruct"
    openrouter_model_gpt_oss: str = "openai/gpt-oss-120b:free"
    openrouter_model_nemotron: str = "nvidia/nemotron-3-nano-30b-a3b:free"
    openrouter_site_url: str = "https://docdoe.ai"
    openrouter_app_name: str = "DocDoe AI"
    ai_max_retries: int = 2
    ai_timeout_seconds: int = 120  # 60s is too short for llama-3.3-70b free tier (~90s p50)
    ai_fallback_to_mock: bool = True
    mock_ai_model_name: str = "mock-exam-tutor-v1"
    # --- Cost-control provider router ---
    ai_router_enabled: bool = False
    ai_router_default_tier: str = "cheap"
    ai_router_daily_budget_usd: float = 0
    ai_router_allow_free_providers: bool = True
    ai_router_provider_daily_units_default: int = 0
    generation_cache_prompt_version: str = "v1"
    # --- Legacy Gemini fields (kept so old .env files don't crash) ---
    gemini_api_key: str | None = Field(
        default=None,
        validation_alias=AliasChoices("GEMINI_API_KEY", "GEMINI_API"),
    )
    gemini_model_flash: str = "gemini-1.5-flash"
    gemini_model_pro: str = "gemini-1.5-pro"
    # --- Additional router provider keys/config ---
    groq_api_key: str | None = Field(default=None, validation_alias=AliasChoices("GROQ_API_KEY"))
    groq_base_url: str = "https://api.groq.com/openai/v1"
    groq_model_main: str = "llama-3.1-8b-instant"
    # --- Z.AI / GLM (free text + vision Flash models) ---
    # OpenAI-compatible: https://api.z.ai/api/paas/v4/
    # Free: glm-4.7-flash (text), glm-4.6v-flash (vision / image understanding).
    zai_api_key: str | None = Field(
        default=None,
        validation_alias=AliasChoices("ZAI_API_KEY", "Z_AI_API_KEY", "GLM_API_KEY"),
    )
    zai_base_url: str = Field(
        default="https://api.z.ai/api/paas/v4",
        validation_alias=AliasChoices("ZAI_BASE_URL", "Z_AI_BASE_URL"),
    )
    zai_model_text: str = Field(
        default="glm-4.7-flash",
        validation_alias=AliasChoices("ZAI_MODEL_TEXT", "GLM_MODEL_TEXT"),
    )
    zai_model_vision: str = Field(
        default="glm-4.6v-flash",
        validation_alias=AliasChoices("ZAI_MODEL_VISION", "GLM_MODEL_VISION"),
    )
    zai_timeout_seconds: int = Field(
        default=90,
        validation_alias=AliasChoices("ZAI_TIMEOUT_SECONDS"),
    )
    nvidia_nim_api_key: str | None = Field(default=None, validation_alias=AliasChoices("NVIDIA_NIM_API_KEY"))
    nvidia_nim_base_url: str = "https://integrate.api.nvidia.com/v1"
    nvidia_nim_model_main: str = "meta/llama-3.1-8b-instruct"  # fast (~3-4s), good for quick answers
    # When true, fast tasks (notes/simple/quiz/flashcards) try NVIDIA NIM first
    # for ~3-4s answers; deep tasks (exam/pyq/video) still use Sarvam.
    fast_provider_enabled: bool = Field(default=False, validation_alias=AliasChoices("FAST_PROVIDER_ENABLED"))
    cloudflare_account_id: str | None = Field(default=None, validation_alias=AliasChoices("CLOUDFLARE_ACCOUNT_ID"))
    cloudflare_api_token: str | None = Field(default=None, validation_alias=AliasChoices("CLOUDFLARE_API_TOKEN"))
    cloudflare_workers_ai_text_model: str = "@cf/meta/llama-3.1-8b-instruct"
    cloudflare_workers_ai_image_model: str = "@cf/black-forest-labs/flux-1-schnell"
    huggingface_api_key: str | None = Field(default=None, validation_alias=AliasChoices("HUGGINGFACE_API_KEY"))
    huggingface_text_model: str = "mistralai/Mistral-7B-Instruct-v0.3"
    huggingface_image_model: str = "stabilityai/stable-diffusion-xl-base-1.0"
    replicate_api_token: str | None = Field(default=None, validation_alias=AliasChoices("REPLICATE_API_TOKEN"))
    # Deepgram powers Record Lecture (speech-to-text). Absent key => the
    # transcription route returns an honest "not configured" result instead of
    # failing, exactly like the other optional providers.
    deepgram_api_key: str | None = Field(default=None, validation_alias=AliasChoices("DEEPGRAM_API_KEY"))
    deepgram_model: str = "nova-2"
    # AI4Bharat Indic Parler is the default teacher voice: it supports Indian
    # English and Malayalam in one local, commercially permissive model.
    tts_provider: str = "ai4bharat"
    tts_default_provider: str = "ai4bharat"
    video_tts_provider: str = "ai4bharat"
    tts_default_voice: str = "nila"
    tts_output_dir: str = "public/generated/audio"
    audio_scene_timeout_seconds: int = 45
    audio_total_timeout_seconds: int = 600
    audio_max_retries: int = 2
    audio_max_concurrency: int = 5
    kokoro_voice: str = "af_heart"
    kokoro_speed: float = 0.95
    ai4bharat_tts_model: str = "ai4bharat/indic-parler-tts"
    ai4bharat_tts_speaker: str = "Anjali"
    ai4bharat_tts_english_speaker: str = "Mary"
    ai4bharat_tts_device: str = "auto"
    # Full precision is validated on the RTX 3060 12 GB and is materially
    # clearer for Malayalam than half precision with Indic Parler-TTS.
    ai4bharat_tts_precision: str = "float32"
    ai4bharat_tts_description: str = (
        "Mary speaks like a warm and patient Kerala secondary-school teacher at a "
        "slightly slow conversational pace. Her voice is clear, close-sounding, "
        "slightly expressive, and recorded with very clear audio and no background noise."
    )
    ai4bharat_tts_malayalam_description: str = (
        "Anjali speaks with a warm, clear and neutral Kerala teacher voice at a "
        "normal calm pace in a close-sounding environment. Her recording has "
        "excellent, very clear audio with no background noise."
    )
    indic_tts_model: str | None = None
    indic_tts_language: str = "ml"
    # Default English voice tuned for natural teaching cadence.
    # AvaMultilingualNeural is a warmer, slower-paced US neural voice that
    # carries explanation content far better than the older Indian newsreader.
    # Negative rate gives space between sentences so concepts land.
    edge_tts_voice_en: str = "en-US-AvaMultilingualNeural"
    edge_tts_voice_ml: str = "ml-IN-SobhanaNeural"
    edge_tts_rate: str = "-4%"
    elevenlabs_api_key: str | None = None
    elevenlabs_voice_id: str | None = None
    audio_normalize: bool = True
    # Distributed state + observability (env-gated; in-memory/no-op when unset)
    redis_url: str | None = None
    sentry_dsn: str | None = None
    sentry_traces_sample_rate: float = 0.1
    email_provider: str = "disabled"
    resend_api_key: str | None = Field(
        default=None,
        validation_alias=AliasChoices("RESEND_API_KEY"),
    )
    resend_base_url: str = "https://api.resend.com"
    email_from: str = "DocDoe <support@docdoe.in>"
    password_reset_token_minutes: int = 30
    password_reset_request_cooldown_seconds: int = 60
    generated_video_jobs_dir: str = "generated-video-jobs"
    generated_video_output_dir: str = "generated-videos"
    video_render_timeout_seconds: int = 3600
    video_min_output_bytes: int = 4096
    video_beta_max_scenes: int = 40
    video_beta_max_duration_seconds: int = 1800
    video_scene_text_max_chars: int = 700
    video_max_concurrent_render_jobs_per_user: int = 1
    video_stuck_job_timeout_minutes: int = 30
    video_enforce_beta_limits_in_development: bool = False
    storage_provider: str = "local"
    storage_bucket: str | None = None
    storage_region: str | None = None
    storage_endpoint_url: str | None = None
    storage_access_key_id: str | None = None
    storage_secret_access_key: str | None = None
    storage_public_base_url: str | None = None
    storage_upload_prefix: str = "uploads"
    storage_audio_prefix: str = "generated-audio"
    storage_video_prefix: str = "generated-videos"
    # Cloudinary (used when storage_provider == "cloudinary")
    cloudinary_cloud_name: str | None = None
    cloudinary_api_key: str | None = None
    cloudinary_api_secret: str | None = None
    auth_provider: str = "jwt"
    # Local preview is an explicit opt-in so a non-production deployment cannot
    # become a shared demo account by accident.
    auth_enabled: bool = True
    allow_insecure_dev_auth: bool = False
    frontend_base_url: str = "http://127.0.0.1:3000"
    google_client_id: str | None = None
    google_client_secret: str | None = None
    google_oauth_redirect_uri: str | None = None
    supabase_url: str | None = None
    supabase_jwt_secret: str | None = None
    supabase_anon_key: str | None = None
    # Backend-only key used for destructive Supabase Auth administration, such
    # as honoring an authenticated account-deletion request. Never expose this
    # value through a NEXT_PUBLIC_* variable or a frontend response.
    supabase_service_role_key: str | None = Field(
        default=None,
        validation_alias=AliasChoices("SUPABASE_SERVICE_ROLE_KEY"),
    )
    jwt_secret_key: str | None = None
    jwt_algorithm: str = "HS256"
    access_token_expire_minutes: int = 10080  # 7 days — keeps users logged in across sessions
    # --- Beta invite gate ---
    beta_access_enabled: bool = False
    beta_invite_code: str | None = None
    # --- Per-user AI rate limiting ---
    rate_limit_enabled: bool = True
    rate_limit_ai_requests_per_day: int = 50
    # --- Admin access ---
    admin_emails: str = ""  # comma-separated list of admin email addresses
    # --- PYQ discovery (official-source-first, safe-disabled by default) ---
    pyq_discovery_enabled: bool = False
    pyq_search_provider: str = "disabled"  # disabled | google_cse | bing | serpapi | manual
    pyq_search_api_key: str | None = Field(
        default=None,
        validation_alias=AliasChoices("PYQ_SEARCH_API_KEY"),
    )
    pyq_search_engine_id: str | None = Field(
        default=None,
        validation_alias=AliasChoices("PYQ_SEARCH_ENGINE_ID"),
    )
    pyq_discovery_max_candidates: int = 12
    pyq_discovery_min_confidence_for_auto_import: float = 0.85

    # --- Exa web search (explicit/current-information questions only) ---
    # The key is read by the backend only. It must never be exposed through a
    # NEXT_PUBLIC_* variable or returned by an API response.
    exa_api_key: str | None = Field(
        default=None,
        validation_alias=AliasChoices("EXA_API_KEY"),
    )
    exa_base_url: str = "https://api.exa.ai"
    exa_search_timeout_seconds: float = 15.0
    exa_search_max_results: int = 6

    # --- Stripe Billing (real SaaS upgrades) ---
    stripe_secret_key: str | None = Field(
        default=None, validation_alias=AliasChoices("STRIPE_SECRET_KEY")
    )
    stripe_webhook_secret: str | None = Field(
        default=None, validation_alias=AliasChoices("STRIPE_WEBHOOK_SECRET")
    )
    stripe_api_version: str = "2026-06-24.dahlia"
    # Stable integration label registered for this checkout flow. The suffix
    # was generated once (rather than per request) so Stripe idempotency keys
    # always see identical request parameters.
    stripe_checkout_integration_identifier: str = "docdoe_web_psuqtgkx"
    # Map the two student-facing subscription keys to recurring Stripe Prices.
    stripe_price_ids: str = ""  # e.g. popular_299=price_xxx,premium_599=price_yyy

    model_config = SettingsConfigDict(
        env_file=BACKEND_DIR / ".env",
        env_file_encoding="utf-8",
        extra="ignore",
    )

    @property
    def is_production(self) -> bool:
        """True when ENVIRONMENT is production (case-insensitive)."""
        return self.environment.strip().casefold() == "production"

    @property
    def cors_origin_list(self) -> list[str]:
        origins = [origin.strip() for origin in self.cors_origins.split(",") if origin.strip()]
        # SECURITY: with allow_credentials=True a "*" makes Starlette reflect ANY
        # origin (echoing the caller's Origin header), so any website could make
        # authenticated requests. Never allow a wildcard in production — drop it
        # and fall back to the known app domains if nothing valid remains.
        if self.environment == "production":
            origins = [o for o in origins if o != "*" and not o.startswith("http://")]
            if not origins:
                origins = ["https://www.docdoe.in", "https://docdoe.in"]
        return origins

    @property
    def resolved_upload_dir(self) -> Path:
        upload_path = Path(self.upload_dir)
        if upload_path.is_absolute():
            return upload_path
        return BACKEND_DIR / upload_path

    @property
    def resolved_tts_output_dir(self) -> Path:
        output_path = Path(self.tts_output_dir)
        if output_path.is_absolute():
            return output_path
        return PROJECT_ROOT / output_path

    @property
    def stripe_enabled(self) -> bool:
        return bool(self.stripe_secret_key)

    def get_stripe_price_id(self, plan_key: str) -> str | None:
        raw = (self.stripe_price_ids or "").strip()
        if not raw:
            return None
        for pair in raw.split(","):
            if "=" in pair:
                k, v = pair.split("=", 1)
                if k.strip().lower() == plan_key.lower():
                    return v.strip() or None
        return None

    @property
    def resolved_generated_video_jobs_dir(self) -> Path:
        jobs_path = Path(self.generated_video_jobs_dir)
        if jobs_path.is_absolute():
            return jobs_path
        return PROJECT_ROOT / jobs_path

    @property
    def resolved_generated_video_output_dir(self) -> Path:
        output_path = Path(self.generated_video_output_dir)
        if output_path.is_absolute():
            return output_path
        return PROJECT_ROOT / output_path


@lru_cache
def get_settings() -> Settings:
    return Settings()