<?php
namespace Clusterlane\Core;

if ( ! defined( 'ABSPATH' ) ) exit;

/**
 * Relevance — how naturally one article could link to another, without AI.
 *
 * Clusters say two articles share a theme; this says whether they actually
 * talk about the same things. Three cheap signals, all from the database:
 *
 *   - word overlap   title (×3), tags (×2) and opening text (×1), cosine-style
 *   - mention        the source already names the target's subject
 *                    (e.g. "Alvik" in a hotel round-up → link to "Hotels in Alvik")
 *   - shared tags    same WordPress tags
 *
 * score() returns 0–100 plus short, human reasons for the UI.
 */
final class Relevance {

    /** @var array<int, array{w: array<string,float>, text: string, tags: string[]}> */
    private static array $cache = [];

    private const STOP = [
        // English
        'the','and','for','with','from','your','you','this','that','are','was','were','has','have','had','its','our','their',
        'into','over','more','most','best','guide','what','when','where','which','how','why','about','all','any','can','will',
        'visit','visiting','things','top','also','just','than','then','them','they','there','here','one','two','new','near',
        // Swedish
        'och','för','med','som','det','den','att','till','från','på','av','är','var','har','inte','ett','en','om','eller',
        'men','kan','ska','vid','över','efter','under','mot','alla','sin','sina','dess','denna','detta','dessa',
        'här','där','när','hur','vad','vem','sedan','också','bara','mycket','många','finns','blir','kommer','även','eller','samt','inom','mellan','utan','sig','man','den','det','de','vi','ni','han','hon',
        'would','could','should','into','onto','very','much','many','some','such','only','other','there','these','those','where',
    ];

    /** Weighted word profile of a post (cached per request). */
    private static function profile( int $post_id ): array {
        if ( isset( self::$cache[ $post_id ] ) ) return self::$cache[ $post_id ];

        $post = get_post( $post_id );
        if ( ! $post ) return self::$cache[ $post_id ] = [ 'w' => [], 'text' => '', 'tags' => [] ];

        $plain = wp_strip_all_tags( strip_shortcodes( $post->post_content ) );
        $plain = html_entity_decode( $plain, ENT_QUOTES | ENT_HTML5, 'UTF-8' );
        $open  = implode( ' ', array_slice( preg_split( '/\s+/u', $post->post_excerpt . ' ' . $plain ), 0, 120 ) );
        $title = html_entity_decode( $post->post_title, ENT_QUOTES | ENT_HTML5, 'UTF-8' );
        $tags  = get_the_terms( $post_id, 'post_tag' ); // cached (wp_get_post_terms() queries every time)
        $tags  = ( ! $tags || is_wp_error( $tags ) ) ? [] : array_map( fn( $t ) => html_entity_decode( $t->name, ENT_QUOTES, 'UTF-8' ), $tags );

        $w = [];
        foreach ( [ [ $title, 3.0 ], [ implode( ' ', $tags ), 2.0 ], [ $open, 1.0 ] ] as [ $txt, $weight ] ) {
            foreach ( self::words( $txt ) as $word ) {
                $w[ $word ] = ( $w[ $word ] ?? 0 ) + $weight;
            }
        }
        return self::$cache[ $post_id ] = [ 'w' => $w, 'text' => mb_strtolower( $title . ' ' . $plain, 'UTF-8' ), 'tags' => $tags ];
    }

    /** @return string[] */
    private static function words( string $txt ): array {
        preg_match_all( '/[\p{L}\p{N}]{3,}/u', mb_strtolower( $txt, 'UTF-8' ), $m );
        return array_values( array_filter( $m[0], fn( $w ) => ! in_array( $w, self::STOP, true ) && ! ctype_digit( $w ) ) );
    }

    /** @var array<string,int>|null word → number of site titles containing it */
    private static ?array $df = null;
    private static int $titles = 0;

    /** How many published titles contain each word (once per request). */
    private static function df( string $word ): int {
        if ( self::$df === null ) {
            global $wpdb;
            // phpcs:disable PluginCheck.Security.DirectDB.UnescapedDBParameter, WordPress.DB.PreparedSQL.NotPrepared -- SQL assembled from code-only fragments (no values); every value is a placeholder passed to $wpdb->prepare().
            $rows = $wpdb->get_col( $wpdb->prepare( 'SELECT post_title FROM %i WHERE ' . Scope::type_sql() . ' AND ' . Scope::sql(), $wpdb->posts ) ); // phpcs:ignore WordPress.DB.DirectDatabaseQuery, WordPress.DB.PreparedSQL.NotPrepared -- titles once per request; type_sql() is prepared, sql() has no values
            // phpcs:enable PluginCheck.Security.DirectDB.UnescapedDBParameter, WordPress.DB.PreparedSQL.NotPrepared
            self::$titles = count( $rows );
            self::$df = [];
            foreach ( $rows as $t ) {
                foreach ( array_unique( self::words( html_entity_decode( $t, ENT_QUOTES | ENT_HTML5, 'UTF-8' ) ) ) as $w ) {
                    self::$df[ $w ] = ( self::$df[ $w ] ?? 0 ) + 1;
                }
            }
        }
        return self::$df[ $word ] ?? 0;
    }

    /** A word is distinctive when few titles on the site use it ("Alvik" yes, "Hotels"/"Stockholm" no). */
    private static function distinctive( string $word ): bool {
        self::df( $word );
        return self::df( $word ) <= max( 3, (int) ceil( self::$titles * 0.04 ) );
    }

    /**
     * The target's distinctive name words, as written in its title
     * ("Hotels in Alvik: Best Stays…" → ["Alvik"]).
     *
     * @return string[]
     */
    private static function subject_words( int $post_id ): array {
        $t = html_entity_decode( get_the_title( $post_id ), ENT_QUOTES | ENT_HTML5, 'UTF-8' );
        $t = trim( preg_split( '/\s*[:–—|(]\s*|\s+-\s+/u', $t )[0] ?? $t );
        preg_match_all( '/[\p{L}\p{N}][\p{L}\p{N}’\'-]{2,}/u', $t, $m );
        $out = [];
        foreach ( $m[0] as $w ) {
            $lw = mb_strtolower( $w, 'UTF-8' );
            if ( in_array( $lw, self::STOP, true ) || mb_strlen( $lw ) < 4 ) continue;
            if ( self::distinctive( $lw ) ) $out[] = $w;
        }
        return $out;
    }

    /**
     * How well $source fits as a place to link to $target.
     *
     * @return array{score:int, reasons:string[]}
     */
    public static function score( int $source_id, int $target_id ): array {
        $a = self::profile( $source_id );
        $b = self::profile( $target_id );
        $reasons = [];

        // Cosine-style overlap of weighted word profiles.
        $dot = 0.0; $shared = [];
        foreach ( $b['w'] as $word => $wb ) {
            if ( isset( $a['w'][ $word ] ) ) { $dot += min( $a['w'][ $word ], $wb ); $shared[ $word ] = $a['w'][ $word ] * $wb; }
        }
        $na = array_sum( $a['w'] ); $nb = array_sum( $b['w'] );
        $overlap = ( $na > 0 && $nb > 0 ) ? $dot / sqrt( $na * $nb ) : 0.0;   // ~0–1

        // Does the source already name the target's subject (its distinctive words)?
        $named = [];
        foreach ( self::subject_words( $target_id ) as $w ) {
            if ( preg_match( '/(?<![\p{L}\p{N}])' . preg_quote( mb_strtolower( $w, 'UTF-8' ), '/' ) . '(?![\p{L}\p{N}])/u', $a['text'] ) ) $named[] = $w;
        }
        $mentions = (bool) $named;
        if ( $mentions ) $reasons[] = sprintf( /* translators: %s: text */ __( 'Mentions %s', 'clusterlane' ), '“' . implode( '”, “', array_slice( $named, 0, 2 ) ) . '”' );

        $tags = array_values( array_intersect( array_map( 'mb_strtolower', $a['tags'] ), array_map( 'mb_strtolower', $b['tags'] ) ) );
        if ( $tags ) $reasons[] = sprintf( /* translators: %s: text */ _n( 'Shared tag: %s', 'Shared tags: %s', count( $tags ), 'clusterlane' ), implode( ', ', array_slice( $tags, 0, 3 ) ) );

        if ( ! $mentions && $shared ) {
            arsort( $shared );
            $top = array_slice( array_keys( array_filter( $shared, fn( $v, $w ) => self::distinctive( $w ), ARRAY_FILTER_USE_BOTH ) ), 0, 3 );
            if ( $top ) $reasons[] = sprintf( /* translators: %s: text */ __( 'Both about %s', 'clusterlane' ), implode( ', ', $top ) );
        }

        $score = min( 1.0, $overlap * 2.2 ) * 60 + ( $mentions ? 30 : 0 ) + min( 10, count( $tags ) * 5 );
        return [ 'score' => (int) round( min( 100, $score ) ), 'reasons' => $reasons ];
    }
}