ai-media-hub/backend/services/ranker.go

package services

import (
	"fmt"
	"math/rand"
	"sort"
	"strings"
	"time"
)

const GeminiFallbackReason = "Gemini Vision 응답이 부족해 키워드 기준으로 보강된 결과입니다."

type GeminiBatchStats struct {
	CandidateCap      int      `json:"candidateCap"`
	Requested         int      `json:"requested"`
	Batches           int      `json:"batches"`
	Succeeded         int      `json:"succeeded"`
	Failed            int      `json:"failed"`
	SequentialRetried int      `json:"sequentialRetried"`
	RecommendedCount  int      `json:"recommendedCount"`
	Errors            []string `json:"errors,omitempty"`
}

func RankSearchResults(query string, results []SearchResult) []SearchResult {
	queryTerms := strings.Fields(strings.ToLower(query))
	positiveTerms := []string{
		"b-roll", "b roll", "stock", "stock footage", "footage", "cinematic", "editorial",
		"establishing", "4k", "hd", "drone", "ambient", "scene", "urban", "cityscape",
	}
	negativeTerms := []string{
		"shocking", "amazing", "crazy", "must watch", "reaction", "gossip", "celebrity",
		"thumbnail", "meme", "prank", "drama", "breaking", "viral", "tutorial",
		"how to", "review", "walkthrough", "course", "lesson", "podcast", "interview",
		"premiere pro", "after effects", "explained", "breakdown", "vlog",
	}
	type scoredResult struct {
		item  SearchResult
		score int
	}

	scored := make([]scoredResult, 0, len(results))
	for _, result := range results {
		score := 0
		text := strings.ToLower(result.Title + " " + result.Snippet + " " + result.Source)
		for _, term := range queryTerms {
			if strings.Contains(text, term) {
				score += 3
			}
		}
		for _, term := range positiveTerms {
			if strings.Contains(text, term) {
				score += 2
			}
		}
		for _, term := range negativeTerms {
			if strings.Contains(text, term) {
				score -= 4
			}
		}
		if result.ThumbnailURL != "" {
			score += 2
		}
		if result.PreviewVideoURL != "" {
			score += 3
		}
		switch result.Source {
		case "Google Video":
			score -= 1
		case "Envato":
			score += 7
		case "Artgrid":
			score += 7
		}
		scored = append(scored, scoredResult{item: result, score: score})
	}

	sort.SliceStable(scored, func(i, j int) bool {
		return scored[i].score > scored[j].score
	})

	ranked := make([]SearchResult, 0, len(scored))
	for _, item := range scored {
		ranked = append(ranked, item.item)
	}
	return ranked
}

func GeminiCandidateLimit(total int) int {
	return total
}

func EvaluateAllCandidatesWithGemini(service *GeminiService, query string, ranked []SearchResult) ([]AIRecommendation, GeminiBatchStats, error) {
	if service == nil {
		return nil, GeminiBatchStats{}, fmt.Errorf("gemini service is not configured")
	}

	limit := GeminiCandidateLimit(len(ranked))
	stats := GeminiBatchStats{
		CandidateCap: limit,
		Requested:    min(limit, len(ranked)),
	}
	stats.Batches = limit
	if limit == 0 {
		return []AIRecommendation{}, stats, nil
	}

	merged := make([]AIRecommendation, 0, len(ranked))
	seen := map[string]bool{}
	for idx := 0; idx < limit; idx++ {
		recommendations, err := recoverGeminiCandidateSequentially(service, query, ranked[idx])
		if err != nil {
			stats.Failed++
			if len(stats.Errors) < 5 {
				stats.Errors = append(stats.Errors, err.Error())
			}
			continue
		}
		stats.Succeeded++
		if len(recommendations) == 0 {
			continue
		}
		stats.SequentialRetried++
		for _, item := range recommendations {
			if item.Link == "" || seen[item.Link] {
				continue
			}
			seen[item.Link] = true
			merged = append(merged, item)
		}
	}
	stats.RecommendedCount = len(merged)

	switch {
	case len(merged) > 0 && stats.Failed == 0:
		return merged, stats, nil
	case len(merged) > 0 && stats.Failed > 0:
		return merged, stats, fmt.Errorf("gemini vision partially failed on %d of %d candidates", stats.Failed, stats.Batches)
	case stats.Failed == stats.Batches:
		if len(stats.Errors) > 0 {
			return nil, stats, fmt.Errorf("gemini vision failed for all candidates: %s", strings.Join(stats.Errors, "; "))
		}
		return nil, stats, fmt.Errorf("gemini vision failed for all candidates")
	default:
		return nil, stats, fmt.Errorf("gemini vision returned no candidate evaluations")
	}
}

func BuildFallbackRecommendations(ranked []SearchResult, limit int, reason string) []AIRecommendation {
	if strings.TrimSpace(reason) == "" {
		reason = GeminiFallbackReason
	}

	fallback := make([]AIRecommendation, 0, min(limit, len(ranked)))
	for _, item := range ranked[:min(limit, len(ranked))] {
		fallback = append(fallback, AIRecommendation{
			Title:           item.Title,
			Link:            item.Link,
			Snippet:         item.Snippet,
			ThumbnailURL:    item.ThumbnailURL,
			PreviewVideoURL: item.PreviewVideoURL,
			Source:          item.Source,
			Reason:          reason,
			Recommended:     false,
		})
	}
	return fallback
}

func RandomizeTopRecommendations(items []AIRecommendation, window int) []AIRecommendation {
	if len(items) < 2 || window < 2 {
		return items
	}

	limit := min(window, len(items))
	shuffled := make([]AIRecommendation, len(items))
	copy(shuffled, items)
	rng := rand.New(rand.NewSource(time.Now().UnixNano()))
	rng.Shuffle(limit, func(i, j int) {
		shuffled[i], shuffled[j] = shuffled[j], shuffled[i]
	})
	return shuffled
}

func recoverGeminiCandidateSequentially(service *GeminiService, query string, candidate SearchResult) ([]AIRecommendation, error) {
	var lastErr error
	for attempt := 0; attempt < 3; attempt++ {
		recs, err := service.Recommend(query, []SearchResult{candidate})
		if err != nil {
			lastErr = err
			time.Sleep(450 * time.Millisecond)
			continue
		}
		return recs, nil
	}
	if lastErr == nil {
		lastErr = fmt.Errorf("gemini vision sequential retry returned no result")
	}
	return nil, lastErr
}

func NeedsSupplementalExploration(items []AIRecommendation) bool {
	if len(items) == 0 {
		return true
	}

	recommendedCount := 0
	negativeCount := 0
	for _, item := range items {
		if item.Recommended {
			recommendedCount++
		}
		if looksNegativeReason(item.Reason) {
			negativeCount++
		}
	}
	if recommendedCount >= 5 {
		return false
	}
	return negativeCount >= max(2, len(items)/2)
}

func looksNegativeReason(reason string) bool {
	lower := strings.ToLower(strings.TrimSpace(reason))
	if lower == "" {
		return false
	}
	for _, token := range []string{
		"부적합", "관련이 없", "맞지 않", "의도와 맞지", "무관", "연관성 낮", "적절하지 않", "불일치",
		"not relevant", "irrelevant", "mismatch", "does not match", "unsuitable",
	} {
		if strings.Contains(lower, token) {
			return true
		}
	}
	return false
}

func MergeRecommendations(recommended []AIRecommendation, ranked []SearchResult, limit int) []AIRecommendation {
	merged := make([]AIRecommendation, 0, min(limit, len(ranked)))
	seen := map[string]bool{}

	for _, item := range recommended {
		if !item.Recommended || shouldExcludeRecommendation(item) {
			continue
		}
		if item.Link == "" || seen[item.Link] {
			continue
		}
		seen[item.Link] = true
		merged = append(merged, item)
	}

	for _, item := range recommended {
		if item.Recommended || item.Link == "" || seen[item.Link] || len(merged) >= limit || shouldExcludeRecommendation(item) {
			continue
		}
		seen[item.Link] = true
		merged = append(merged, item)
	}
	return merged
}

func max(a, b int) int {
	if a > b {
		return a
	}
	return b
}

func shouldExcludeRecommendation(item AIRecommendation) bool {
	if strings.Contains(item.Reason, GeminiFallbackReason) {
		return true
	}
	return looksNegativeReason(item.Reason)
}