"""Ensemble fusion — combines DetectionResult objects into a final verdict.

Strategy:
  1. Filter: drop failed/errored detectors (set confidence = 0 proxy).
  2. Bayesian log-odds combination weighted by detector weights × confidence.
  3. Conflict detection: flag when high-confidence detectors strongly disagree.
  4. Threshold: map final p_fake → REAL / UNCERTAIN / AI_GENERATED.

Why Bayesian log-odds over simple weighted average:
  - A detector that says 0.99 should dominate more than one saying 0.6.
  - Log-odds is the correct way to combine independent probability estimates.
  - We still weight by (detector_weight × confidence) to account for
    detector reliability and self-reported uncertainty.
"""

from __future__ import annotations

import math
from typing import Sequence

from backend.core.config import settings
from backend.core.schema import DetectionResult, EnsembleResult, Verdict
from backend.fusion.meta_learner import get_meta_learner

# Weight per detector — mirrors config but looked up at fusion time.
_WEIGHTS: dict[str, float] = {
    "hive":             settings.weight_hive,
    "hive_vlm":         settings.weight_hive_vlm,
    "aiornot":          settings.weight_aiornot,
    "sightengine":      settings.weight_sightengine,
    "claude_vision":    settings.weight_claude,
    "mistral_vision":   settings.weight_mistral,
    "gpt4_vision":      settings.weight_gpt4,
    "gemini_vision":    settings.weight_gemini,
    "gemma_vision":       settings.weight_gemma_vision,
    "openrouter_gpt4":    settings.weight_openrouter_gpt4,
    "openrouter_llama":   settings.weight_openrouter_llama,
    "openrouter_mistral": settings.weight_openrouter_mistral,
    "local_forensics":    settings.weight_local_forensics,
    "local_finetuned":    settings.weight_local_finetuned,
    "hybrid_model":       settings.weight_hybrid_model,
    # V3 detectors
    "frequency_domain":   settings.weight_frequency_domain,
    "clip_unifd":         settings.weight_clip_unifd,
    "dire":               settings.weight_dire,
    "face_specialist":    settings.weight_face_specialist,
    # V4 detectors
    "siglip2_aigc":       settings.weight_siglip2_aigc,
    "npr":                settings.weight_npr,
}


# Per-detector vote multiplier for the voting strategy. Sightengine is the
# primary commercial signal, so its vote counts double.
_VOTE_WEIGHTS: dict[str, float] = {"sightengine": 2.0}


def _vote_weight(detector: str) -> float:
    return _VOTE_WEIGHTS.get(detector, 1.0)


class EnsembleFusion:

    def fuse(
        self,
        results: Sequence[DetectionResult],
        processing_time_ms: float = 0.0,
        fusion_strategy_override: str | None = None,
    ) -> EnsembleResult:
        valid = [r for r in results if r.error is None and r.confidence > 0]
        failed = [r for r in results if r.error is not None or r.confidence == 0]

        if not valid:
            return EnsembleResult(
                verdict=Verdict.UNCERTAIN,
                p_fake=0.5,
                confidence=0.0,
                primary_evidence="All detectors failed or were unavailable",
                uncertainty_factors=[f"{r.detector}: {r.error}" for r in failed],
                detector_results=list(results),
                fusion_details={"method": "none"},
                processing_time_ms=processing_time_ms,
            )

        # Choose fusion strategy: meta-learner → per-request override → server default
        active_strategy = fusion_strategy_override or settings.fusion_strategy
        ml = get_meta_learner() if settings.use_meta_learner_fusion else None
        if ml is not None:
            p_fake, method_details = ml.predict(valid)
        elif active_strategy == "voting":
            p_fake, method_details = self._voting_fusion(valid)
        else:
            p_fake, method_details = self._bayesian_fusion(valid)
        verdict = _verdict_from_p_fake(p_fake, detector_results=valid)
        conflicts = self._detect_conflicts(valid)

        raw_conf = abs(p_fake - 0.5) * 2.0
        conflict_penalty = 0.15 * len(conflicts)
        confidence = float(max(0.0, min(1.0, raw_conf - conflict_penalty)))

        primary_evidence, supporting = self._summarise_evidence(valid, verdict)

        uncertainty_factors = conflicts[:]
        if failed:
            uncertainty_factors.append(
                f"Unavailable detectors: {', '.join(r.detector for r in failed)}"
            )

        # Extract diffusion_suspicion from local_forensics raw output
        diffusion_suspicion = 0.0
        for r in results:
            if r.detector == "local_forensics" and r.raw:
                diffusion_suspicion = float(r.raw.get("diffusion_suspicion", 0.0))
                break

        # Propagate generator name from whichever detector identified it
        generator = next(
            (r.generator for r in results if r.generator),
            None,
        )

        return EnsembleResult(
            verdict=verdict,
            p_fake=p_fake,
            confidence=confidence,
            primary_evidence=primary_evidence,
            supporting_evidence=supporting,
            uncertainty_factors=uncertainty_factors,
            detector_results=list(results),
            fusion_details={
                "method": "meta_learner" if ml else active_strategy,
                **method_details,
                "n_valid": len(valid),
                "n_failed": len(failed),
            },
            processing_time_ms=processing_time_ms,
            diffusion_suspicion=diffusion_suspicion,
            generator=generator,
        )

    # ── Bayesian log-odds fusion ────────────────────────────────

    def _bayesian_fusion(
        self, valid: list[DetectionResult]
    ) -> tuple[float, dict]:
        """
        Combine independent P(fake) estimates via log-odds.
        log_odds_posterior = log_odds_prior + Σ w_i * log_odds_i
        Prior: 0.5 (no prior knowledge).
        """
        prior_log_odds = 0.0

        total_weight = 0.0
        weighted_log_odds = 0.0
        per_detector: dict[str, dict] = {}

        for r in valid:
            w = _WEIGHTS.get(r.detector, 1.0) * float(r.confidence)

            # Dynamic weight adjustment based on generator detection
            # If Sightengine detected a specific generator, boost its weight (30% boost)
            if r.detector == "sightengine" and hasattr(r, "generator") and r.generator:
                w *= 1.3  # 30% confidence boost when generator explicitly detected
                weight_reason = f"generator detected ({r.generator})"
            else:
                weight_reason = "base weight"

            p = float(r.p_fake)
            p = max(0.01, min(0.99, p))
            lo = math.log(p / (1.0 - p))

            weighted_log_odds += w * lo
            total_weight += w
            per_detector[r.detector] = {
                "p_fake": r.p_fake,
                "confidence": r.confidence,
                "weight": _WEIGHTS.get(r.detector, 1.0),
                "effective_weight": w,
                "log_odds": lo,
                "weight_reason": weight_reason,
                "generator": getattr(r, "generator", None),
            }

        if total_weight == 0:
            return 0.5, {}

        avg_log_odds = prior_log_odds + (weighted_log_odds / total_weight)
        p_fake = 1.0 / (1.0 + math.exp(-avg_log_odds))

        return float(p_fake), {
            "total_weight": total_weight,
            "avg_log_odds": avg_log_odds,
            "per_detector": per_detector,
        }

    # ── Voting fusion ───────────────────────────────────────────

    def _voting_fusion(
        self, valid: list[DetectionResult]
    ) -> tuple[float, dict]:
        """
        Each detector casts one vote based on its own verdict threshold.
        p_fake is computed as fraction of AI votes, then sharpened so it
        actually crosses the ai_threshold / real_threshold.

        voting_threshold (default 0.5): fraction of votes needed to call AI.
        """
        votes_ai = [r for r in valid if r.p_fake >= settings.ai_threshold]
        votes_real = [r for r in valid if r.p_fake <= settings.real_threshold]
        votes_uncertain = [
            r for r in valid
            if settings.real_threshold < r.p_fake < settings.ai_threshold
        ]

        # Votes are weighted per detector (Sightengine counts double).
        total_votes = sum(_vote_weight(r.detector) for r in valid)
        ai_votes = sum(_vote_weight(r.detector) for r in votes_ai)
        real_votes = sum(_vote_weight(r.detector) for r in votes_real)
        ai_fraction = ai_votes / total_votes

        # Map fraction → p_fake in a way that crosses thresholds cleanly:
        # If ai_fraction >= voting_threshold → p_fake just above ai_threshold
        # If ai_fraction == 0 → p_fake just below real_threshold
        # Otherwise → UNCERTAIN midpoint
        if ai_fraction >= settings.voting_threshold:
            # Scale within [ai_threshold, 1.0] based on how overwhelming the vote is
            p_fake = settings.ai_threshold + (1.0 - settings.ai_threshold) * ai_fraction
        elif real_votes / total_votes >= settings.voting_threshold:
            real_fraction = real_votes / total_votes
            p_fake = settings.real_threshold * (1.0 - real_fraction)
        else:
            # Split vote → UNCERTAIN: vote-weighted average of individual p_fakes
            p_fake = sum(_vote_weight(r.detector) * r.p_fake for r in valid) / total_votes

        return float(p_fake), {
            "votes_ai": [r.detector for r in votes_ai],
            "votes_real": [r.detector for r in votes_real],
            "votes_uncertain": [r.detector for r in votes_uncertain],
            "ai_fraction": ai_fraction,
            "voting_threshold": settings.voting_threshold,
        }

    # ── Conflict detection ──────────────────────────────────────

    def _detect_conflicts(self, valid: list[DetectionResult]) -> list[str]:
        """Flag when high-confidence detectors strongly disagree.

        Also flags significant disagreement between hybrid model versions
        (latest vs backup) which may indicate uncertain/ambiguous images.
        """
        conflicts: list[str] = []

        high_conf = [r for r in valid if r.confidence >= 0.6]
        if len(high_conf) < 2:
            return conflicts

        ai_detectors   = [r for r in high_conf if r.verdict == Verdict.AI_GENERATED]
        real_detectors = [r for r in high_conf if r.verdict == Verdict.REAL]

        if ai_detectors and real_detectors:
            ai_names   = ", ".join(r.detector for r in ai_detectors)
            real_names = ", ".join(r.detector for r in real_detectors)
            conflicts.append(
                f"Detector disagreement: [{ai_names}] → AI vs [{real_names}] → REAL"
            )

        # Check for hybrid model version disagreement (both are "hybrid_model" detector)
        # This catches cases where latest and backup models give conflicting signals
        hybrid_results = [r for r in valid if r.detector == "hybrid_model"]
        if len(hybrid_results) >= 2:
            # Compare p_fake values across model versions
            p_fakes = [float(r.p_fake) for r in hybrid_results]
            p_fake_spread = max(p_fakes) - min(p_fakes)

            # If models differ by >0.25 in p_fake, flag as uncertainty signal
            # e.g., latest=0.48, backup=0.34 → spread=0.14 (not flagged)
            # e.g., latest=0.48, backup=0.65 → spread=0.17 (not flagged, both >0.3)
            # But if verdicts differ (one AI, one REAL), flag even with smaller spread
            verdicts = [r.verdict for r in hybrid_results]
            if len(set(verdicts)) > 1:  # Different verdicts
                conflicts.append(
                    f"Hybrid model version disagreement: "
                    f"{hybrid_results[0].detector} versions give conflicting signals "
                    f"(p_fake range: {min(p_fakes):.2f}-{max(p_fakes):.2f})"
                )

        return conflicts

    # ── Evidence summary ────────────────────────────────────────

    def _summarise_evidence(
        self, valid: list[DetectionResult], verdict: Verdict
    ) -> tuple[str, list[str]]:
        if verdict == Verdict.AI_GENERATED:
            ranked = sorted(valid, key=lambda r: r.p_fake, reverse=True)
        elif verdict == Verdict.REAL:
            ranked = sorted(valid, key=lambda r: r.p_fake)
        else:
            ranked = sorted(valid, key=lambda r: r.confidence, reverse=True)

        primary = ""
        supporting: list[str] = []

        for r in ranked:
            if not r.evidence:
                continue
            first = r.evidence[0]

            # Make verdict more explicit in evidence summary
            if verdict == Verdict.REAL and r.verdict == Verdict.REAL:
                summary = f"[{r.detector}] Real image: {(1-r.p_fake)*100:.1f}% confidence"
            elif verdict == Verdict.AI_GENERATED and r.verdict == Verdict.AI_GENERATED:
                summary = f"[{r.detector}] AI-generated: {r.p_fake*100:.1f}% confidence"
            elif verdict == Verdict.MANIPULATED_DEEPFAKE and r.verdict == Verdict.MANIPULATED_DEEPFAKE:
                summary = f"[{r.detector}] Face manipulation detected: {r.p_fake*100:.1f}% confidence"
            else:
                summary = f"[{r.detector}] {first}"

            if not primary:
                primary = summary
            else:
                supporting.append(summary)
            if len(supporting) >= 4:
                break

        if not primary:
            primary = f"Ensemble confidence: {(1-valid[0].p_fake)*100:.1f}% real from {len(valid)} detectors"

        return primary, supporting


def _verdict_from_p_fake(
    p_fake: float,
    detector_results: list | None = None,
) -> Verdict:
    """
    Map ensemble p_fake to a verdict.

    If p_fake crosses the AI threshold AND the hybrid_model detector specifically
    predicted MANIPULATED_DEEPFAKE with high confidence, return that more specific
    verdict instead of the generic AI_GENERATED.
    """
    if p_fake >= settings.ai_threshold:
        # Check if the V1 model specifically identified a deepfake manipulation
        if detector_results:
            hybrid = next(
                (r for r in detector_results
                 if r.detector == "hybrid_model" and r.error is None
                 and r.verdict == Verdict.MANIPULATED_DEEPFAKE
                 and r.confidence >= 0.5),
                None,
            )
            if hybrid is not None:
                return Verdict.MANIPULATED_DEEPFAKE
        return Verdict.AI_GENERATED
    if p_fake <= settings.real_threshold:
        return Verdict.REAL
    return Verdict.UNCERTAIN