package analysis

import (
	"math"
	"sort"
	"strings"

	"go-places-api/internal/domain"
)

// SourceLocal identifies badges derived from the picture's own pixels and
// metadata rather than from a trained model.
const SourceLocal = "local"

// badge is a candidate rule. A rule contributes a badge when its condition
// holds, with a confidence derived from how far past the threshold the signal
// sits.
type badge struct {
	Label   string
	Display string
	// Condition returns a confidence in 0..1, or a negative value to reject.
	Condition func(s Signals, md Metadata) float64
}

// DeriveBadges turns measurements into badges.
//
// These are honest heuristics over colour, brightness, detail and EXIF. They
// recognise scene and subject *categories* such as sky, vegetation or a sunset,
// not arbitrary objects; object-level recognition belongs behind the remote
// analyzer, whose labels are merged in by Combine.
func DeriveBadges(res *Result) []domain.Badge {
	if res == nil {
		return nil
	}
	rules := []badge{
		{labelSky, "Sky", ruleSky},
		{labelVegetation, "Vegetation", ruleVegetation},
		{labelNature, "Nature", ruleNature},
		{labelWater, "Water", ruleWater},
		{labelBeach, "Beach", ruleBeach},
		{labelSunset, "Sunset", ruleSunset},
		{labelSnow, "Snow", ruleSnow},
		{labelIndoor, "Indoor", ruleIndoor},
		{labelPortrait, "Portrait", rulePortrait},
		{labelLandscape, "Landscape", ruleLandscape},
		{labelPanorama, "Panorama", rulePanorama},
		{labelLowLight, "Low light", ruleLowLight},
		{labelNight, "Night", ruleNight},
		{labelOverexposed, "Overexposed", ruleOverexposed},
		{labelMonochrome, "Monochrome", ruleMonochrome},
		{labelVivid, "Vivid", ruleVivid},
		{labelCloseUp, "Close-up", ruleCloseUp},
		{labelGraphic, "Graphic", ruleGraphic},
		{labelScreenshot, "Screenshot or document", ruleScreenshot},
	}

	var out []domain.Badge
	for _, r := range rules {
		conf := r.Condition(res.Signals, res.Metadata)
		if conf <= 0 {
			continue
		}
		out = append(out, domain.Badge{
			Label:      r.Label,
			Display:    r.Display,
			Confidence: round2(clamp01(conf)),
			Source:     SourceLocal,
		})
	}
	return out
}

const (
	labelSky         = "sky"
	labelVegetation  = "vegetation"
	labelNature      = "nature"
	labelWater       = "water"
	labelBeach       = "beach"
	labelSunset      = "sunset"
	labelSnow        = "snow"
	labelIndoor      = "indoor"
	labelPortrait    = "portrait"
	labelLandscape   = "landscape"
	labelPanorama    = "panorama"
	labelLowLight    = "low-light"
	labelNight       = "night"
	labelOverexposed = "overexposed"
	labelMonochrome  = "monochrome"
	labelVivid       = "vivid"
	labelCloseUp     = "close-up"
	labelGraphic     = "graphic"
	labelScreenshot  = "screenshot"
)

// Thresholds are collected here so they can be read, and tuned, in one place.
const (
	skyBlueShare       = 0.40
	greenShare         = 0.25
	waterBlueShare     = 0.22
	waterBottomSmooth  = 0.06
	beachSandShare     = 0.28
	sunsetWarmShare    = 0.18
	snowBrightness     = 0.82
	snowLowSaturation  = 0.12
	lowLightBrightness = 0.20
	nightBrightness    = 0.16
	overexposed        = 0.88
	monochromeSat      = 0.10
	vividSat           = 0.42
	closeUpEdges       = 0.30
	graphicEdges       = 0.38
	graphicSaturation  = 0.28
	screenshotEdges    = 0.22
	screenshotBright   = 0.62
)

// scale maps a signal against a threshold into a confidence that grows with
// the margin, saturating at full confidence once the signal is twice the
// threshold. A signal at exactly the threshold yields 0 and is therefore not
// emitted, which keeps thresholds meaning "at least this much".
func scale(value, threshold float64) float64 {
	if threshold <= 0 {
		return 0
	}
	ratio := value / threshold
	if ratio <= 1 {
		return 0
	}
	// Ease out so confidence approaches 1 asymptotically.
	t := (ratio - 1) / (ratio + 1)
	return 0.5 + 0.5*t*2
}

// scaleBelow is the inverse of scale, for signals where a *low* value is the
// positive signal, such as saturation for a monochrome picture. It returns zero
// at or above the threshold and rises as the value falls towards zero.
//
// Thresholding an inverted signal by subtracting it and feeding the result to
// scale would compare a margin against a threshold and silently emit nothing,
// so the two directions are deliberately given separate helpers.
func scaleBelow(value, threshold float64) float64 {
	if threshold <= 0 || value >= threshold {
		return 0
	}
	// ratio runs 0 at the threshold to 1 at zero.
	ratio := (threshold - value) / threshold
	if ratio <= 0 {
		return 0
	}
	return 0.4 + 0.5*ratio
}

func ruleSky(s Signals, _ Metadata) float64 {
	// Sky sits in the upper band and is blue with little detail.
	blue := scale(s.Top.HueWeights.Blue, skyBlueShare)
	smooth := 1 - scale(s.Top.EdgeDensity, 0.10)
	return 0.75*blue + 0.25*smooth*blue
}

func ruleVegetation(s Signals, _ Metadata) float64 {
	// Vegetation is green and can appear in any band, weighted towards the
	// lower half where ground and trees are.
	overall := scale(s.HueWeights.Green, greenShare)
	lower := scale(s.Bottom.HueWeights.Green, greenShare)
	return 0.5*overall + 0.5*lower
}

func ruleNature(s Signals, _ Metadata) float64 {
	green := s.HueWeights.Green
	blue := s.HueWeights.Blue
	// Nature means the frame is dominated by vegetation or sky and is not a
	// document-like composition.
	if s.EdgeDensity > screenshotEdges {
		return 0
	}
	return 0.6*scale(green, greenShare) + 0.4*scale(blue, skyBlueShare)
}

func ruleWater(s Signals, _ Metadata) float64 {
	// Water is blue in the lower band and has little structure, unlike a sky
	// which occupies the upper band.
	if s.Top.HueWeights.Blue > s.Bottom.HueWeights.Blue {
		return 0
	}
	blue := scale(s.Bottom.HueWeights.Blue, waterBlueShare)
	smooth := 1 - scale(s.Bottom.EdgeDensity, waterBottomSmooth)
	return 0.8*blue + 0.2*smooth*blue
}

func ruleBeach(s Signals, _ Metadata) float64 {
	// A beach is sky above and warm sand below.
	sky := scale(s.Top.HueWeights.Blue, skyBlueShare)
	sand := scale(s.Bottom.HueWeights.Warm, beachSandShare)
	return 0.5 * sky * sand
}

func ruleSunset(s Signals, _ Metadata) float64 {
	warm := scale(s.Middle.HueWeights.Warm, sunsetWarmShare)
	// Sunset keeps some light: a pitch-dark frame is night, not sunset.
	bright := 1 - scale(1-s.Brightness, 0.25)
	return warm * bright
}

func ruleSnow(s Signals, _ Metadata) float64 {
	// Snow is very bright and almost colourless.
	bright := scale(s.Brightness, snowBrightness)
	flat := 1 - scale(s.Saturation, snowLowSaturation)
	return 0.7*bright + 0.3*flat*bright
}

func ruleIndoor(s Signals, _ Metadata) float64 {
	// Indoor shots are shot close with plenty of detail and no sky.
	if s.Top.HueWeights.Blue > skyBlueShare/2 {
		return 0
	}
	detail := scale(s.EdgeDensity, closeUpEdges)
	return detail
}

func rulePortrait(s Signals, _ Metadata) float64 {
	return orientationConfidence(s, 0.5, 0.95)
}

func ruleLandscape(s Signals, _ Metadata) float64 {
	return orientationConfidence(s, 1.06, 2.2)
}

func rulePanorama(s Signals, _ Metadata) float64 {
	return orientationConfidence(s, 2.5, 6.0)
}

// orientationConfidence scores how strongly the aspect ratio sits inside the
// band [min,max]. It returns zero outside the band, one at its centre, and
// falls off towards the edges.
func orientationConfidence(s Signals, min, max float64) float64 {
	if s.Width <= 0 || s.Height <= 0 {
		return 0
	}
	ratio := float64(s.Width) / float64(s.Height)
	if ratio < min || ratio > max {
		return 0
	}
	// The comparison happens in log space so a wide band and a tall band are
	// treated symmetrically.
	halfSpan := math.Log(max/min) / 2
	if halfSpan <= 0 {
		return 0
	}
	drift := abs(math.Log(ratio/math.Sqrt(min*max))) / halfSpan
	return 1 - clamp01(drift)
}

// ruleNight prefers the clock time from EXIF and falls back to brightness alone
// when the picture carries no timestamp.
func ruleNight(s Signals, md Metadata) float64 {
	if md.TakenAt != nil {
		if h := md.TakenAt.UTC().Hour(); h >= 21 || h < 5 {
			return 0.9
		}
		// Daytime capture means the frame is not a night scene even if it is
		// dark, which usually means a window or an interior.
		return 0
	}
	return scaleBelow(s.Brightness, nightBrightness)
}

func ruleLowLight(s Signals, _ Metadata) float64 {
	return scaleBelow(s.Brightness, lowLightBrightness)
}

func ruleOverexposed(s Signals, _ Metadata) float64 {
	return scale(s.Brightness, overexposed)
}

func ruleMonochrome(s Signals, _ Metadata) float64 {
	return scaleBelow(s.Saturation, monochromeSat)
}

func ruleVivid(s Signals, _ Metadata) float64 {
	return scale(s.Saturation, vividSat)
}

func ruleCloseUp(s Signals, _ Metadata) float64 {
	return scale(s.EdgeDensity, closeUpEdges)
}

func ruleGraphic(s Signals, _ Metadata) float64 {
	// Graphics and illustrations have heavy outlines but muted colour.
	edges := scale(s.EdgeDensity, graphicEdges)
	muted := 1 - scale(s.Saturation, graphicSaturation)
	return 0.6*edges + 0.4*muted*edges
}

func ruleScreenshot(s Signals, _ Metadata) float64 {
	// Screenshots and scans are flat, bright and edge-dense because text
	// produces many small high-contrast strokes.
	if s.Brightness < screenshotBright {
		return 0
	}
	edges := scale(s.EdgeDensity, screenshotEdges)
	return edges
}

// Combine merges heuristic badges with model badges. Where both sources report
// the same label the model wins, because a trained detector is more reliable
// than a colour heuristic, but the heuristic confidence is retained as a floor
// so a confident measurement is not discarded.
func Combine(local, model []domain.Badge) []domain.Badge {
	byLabel := map[string]domain.Badge{}
	order := []string{}

	add := func(b domain.Badge) {
		if b.Label == "" {
			return
		}
		if existing, ok := byLabel[b.Label]; ok {
			existing.Confidence = maxf(existing.Confidence, b.Confidence)
			// A model label is what actually identified the object, so it wins
			// the display name and keeps its own source.
			if b.Source != SourceLocal {
				existing.Display = b.Display
				existing.Source = b.Source
			}
			byLabel[b.Label] = existing
			return
		}
		byLabel[b.Label] = b
		order = append(order, b.Label)
	}

	for _, b := range model {
		add(b)
	}
	for _, b := range local {
		add(b)
	}

	out := make([]domain.Badge, 0, len(order))
	for _, label := range order {
		b := byLabel[label]
		if b.Display == "" {
			b.Display = Humanize(label)
		}
		b.Confidence = round2(clamp01(b.Confidence))
		out = append(out, b)
	}
	sort.SliceStable(out, func(i, j int) bool {
		if out[i].Confidence != out[j].Confidence {
			return out[i].Confidence > out[j].Confidence
		}
		return out[i].Label < out[j].Label
	})
	return out
}

// Humanize turns a kebab-case label such as "low-light" into "Low light".
func Humanize(label string) string {
	parts := strings.FieldsFunc(label, func(r rune) bool {
		return r == '-' || r == '_' || r == ' '
	})
	if len(parts) == 0 {
		return label
	}
	for i, p := range parts {
		p = strings.ToLower(p)
		// Only the first word is capitalised, so "low-light" reads "Low light"
		// rather than "Low Light".
		if i == 0 && p != "" {
			p = strings.ToUpper(p[:1]) + p[1:]
		}
		parts[i] = p
	}
	return strings.Join(parts, " ")
}

func round2(v float64) float64 { return math.Round(v*100) / 100 }

func abs(v float64) float64 { return math.Abs(v) }

func maxf(a, b float64) float64 { return math.Max(a, b) }