backend/core/config.py

"""Central configuration — loaded once at startup from .env."""

from __future__ import annotations

from pathlib import Path
from typing import List, Tuple
from pydantic_settings import BaseSettings, SettingsConfigDict

# Project root is three levels up from this file (backend/core/config.py)
_PROJECT_ROOT = Path(__file__).parent.parent.parent


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

    # ── Specialized AI detection APIs ───────────────────────────
    hive_api_key: str = ""
    hive_vlm_secret_key: str = ""   # Hive VLM (Bearer token)
    aiornot_api_key: str = ""
    reality_defender_api_key: str = ""
    # Support multiple Sightengine keys for failover (comma-separated)
    sightengine_api_users: str = ""      # e.g., "user1,user2,user3"
    sightengine_api_secrets: str = ""    # e.g., "secret1,secret2,secret3"
    # Backward compatibility: single key (if provided, converted to list with 1 element)
    sightengine_api_user: str = ""
    sightengine_api_secret: str = ""

    # ── VLM APIs ────────────────────────────────────────────────
    anthropic_api_key: str = ""
    openai_api_key: str = ""
    google_api_key: str = ""
    open_route_api_key: str = ""
    mistral_api_key: str = ""

    # ── Detector toggles ────────────────────────────────────────
    enable_hive: bool = False
    enable_hive_vlm: bool = False
    enable_aiornot: bool = False
    enable_sightengine: bool = True
    enable_reality_defender: bool = False
    enable_claude: bool = False
    enable_gpt4: bool = False
    enable_gemini: bool = False
    enable_local_forensics: bool = False
    enable_hybrid_model: bool = False
    enable_openrouter_gemma: bool = False
    enable_openrouter_gpt4: bool = False
    enable_openrouter_llama: bool = False
    enable_openrouter_mistral: bool = False
    enable_mistral: bool = False          # Mistral direct API
    enable_local_finetuned: bool = True   # Fine-tuned Swin Transformer (92% AI recall)

    # ── V3 detector toggles ─────────────────────────────────────
    enable_frequency_domain: bool = True  # FFT spectrum ResNet-18 (Step 2)
    enable_clip_unifd: bool = True        # CLIP ViT-L/14 linear head (Step 3, needs open_clip)
    enable_dire: bool = True              # Diffusion reconstruction error (Step 4 — vae_proxy on CPU, full_dire on GPU)
    enable_face_specialist: bool = False  # F3-Net face forensics (Step 6)
    enable_siglip2_aigc: bool = True      # SigLIP2-base open-deepfake-detection (Nov 2025)
    enable_npr: bool = True               # NPR ResNet-50 upsampling artifact detector (CVPR 2024)
    use_meta_learner_fusion: bool = False  # XGBoost/LR meta-learner (Step 5; False = Bayesian log-odds fallback)
    # fusion_strategy: "weighted" (Bayesian log-odds) | "voting" (majority vote, each detector 1 vote)
    fusion_strategy: str = "weighted"
    # Voting threshold: fraction of detectors that must vote AI to return AI_GENERATED (default 0.5 = majority)
    voting_threshold: float = 0.5

    # ── V1 HybridForensicsModel paths ───────────────────────────
    # Path to the trained checkpoint file (hybrid_detector.pt)
    hybrid_model_checkpoint: str = str(
        _PROJECT_ROOT.parent / "Deepfake-V1" / "Professional_Integration"
        / "ml" / "checkpoints" / "hybrid_detector.pt"
    )
    # Path to Professional_Integration/ so `import ml` works
    hybrid_model_code_root: str = str(
        _PROJECT_ROOT.parent / "Deepfake-V1" / "Professional_Integration"
    )
    # Model versions: latest and backup
    hybrid_model_latest_checkpoint: str = str(
        _PROJECT_ROOT.parent / "Deepfake-V1" / "Professional_Integration"
        / "ml" / "checkpoints" / "hybrid_detector.pt"
    )
    hybrid_model_backup_checkpoint: str = str(
        _PROJECT_ROOT.parent / "Deepfake-V1" / "Professional_Integration"
        / "ml" / "checkpoints" / "hybrid_detector_v1_backup.pt"
    )
    # Default model version to use (latest or backup)
    hybrid_model_default_version: str = "latest"

    # ── Fusion weights (sum need not equal 1 — normalised at runtime) ──
    # Higher = more trusted. Tune after benchmarking.
    weight_hive: float = 2.5              # purpose-built, high accuracy
    weight_hive_vlm: float = 2.0          # Hive VLM — strong vision reasoning
    weight_aiornot: float = 2.0           # diverse training set
    weight_sightengine: float = 3.0       # primary signal
    weight_mistral: float = 2.2           # direct Mistral API — strong multimodal
    weight_claude: float = 2.2            # Anthropic VLM — strong reasoning
    weight_gpt4: float = 1.5
    weight_gemini: float = 0.4            # low weight: Gemini Vision mislabels its own generated images as real
    weight_gemma_vision: float = 1.3
    weight_openrouter_gpt4: float = 2.0
    weight_openrouter_llama: float = 1.5
    weight_openrouter_mistral: float = 1.4
    weight_local_forensics: float = 0.6   # supporting signal only
    weight_local_finetuned: float = 2.5   # fine-tuned Swin: 92% AI recall on domain data
    weight_hybrid_model: float = 1.5

    # V3 detector weights
    weight_frequency_domain: float = 1.5  # trained ResNet-18 on FFT spectra
    weight_clip_unifd: float = 1.2        # UniFD linear head — can FP on compressed images
    weight_dire: float = 1.5              # DIRE reconstruction error (Step 4)
    weight_face_specialist: float = 0.5   # unreliable on non-manipulation AI images
    weight_siglip2_aigc: float = 2.0      # SigLIP2 Nov 2025 — strong on diffusion + AI fashion
    weight_npr: float = 1.8               # NPR CVPR 2024 — upsampling artifacts, fast CPU

    # ── Confidence-based gating thresholds ──────────────────────
    # Used to skip expensive detectors if cheap ones are already confident
    enable_gating: bool = True
    gating_confidence_threshold: float = 0.6  # confidence needed to skip Tier 2+
    gating_p_fake_margin: float = 0.2  # |p_fake - 0.5| margin for early exit

    # ── Verdict thresholds ──────────────────────────────────────
    # Calibrated for accuracy on modern diffusion images (Apr 20):
    # Lower ai_threshold so fused log-odds actually reaches it;
    # raise real_threshold so weak REAL signals don't mask AI images.
    ai_threshold: float = 0.42     # lowered: Sightengine alone can push past this
    real_threshold: float = 0.28   # real images cluster near 0.02-0.04 so safe to lower
    # between thresholds → UNCERTAIN

    # ── HTTP ────────────────────────────────────────────────────
    request_timeout: float = 20.0  # seconds per API call

    # ── MongoDB ──────────────────────────────────────────────────
    mongodb_url: str = "mongodb://localhost:27017"
    mongodb_database: str = "deepshield"

    @property
    def sightengine_pairs(self) -> List[Tuple[str, str]]:
        """Parse Sightengine credentials into (user, secret) pairs for failover.

        Supports both:
        - Multiple keys: SIGHTENGINE_API_USERS=user1,user2 SIGHTENGINE_API_SECRETS=secret1,secret2
        - Single key (backward compat): SIGHTENGINE_API_USER=user SIGHTENGINE_API_SECRET=secret
        """
        # Try multi-key format first
        if self.sightengine_api_users and self.sightengine_api_secrets:
            users = [u.strip() for u in self.sightengine_api_users.split(",") if u.strip()]
            secrets = [s.strip() for s in self.sightengine_api_secrets.split(",") if s.strip()]
            pairs = list(zip(users, secrets))
            if pairs:
                return pairs

        # Fall back to single-key format
        if self.sightengine_api_user and self.sightengine_api_secret:
            return [(self.sightengine_api_user, self.sightengine_api_secret)]

        return []


settings = Settings()