<?php

declare(strict_types=1);

namespace WordPress\CloudflareWorkersAiProvider\Embeddings;

use WordPress\CloudflareWorkersAiProvider\Backfill\Backfill;

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

/**
 * A post (or image) as plain text for embedding, plus the metadata stored with it.
 *
 * Works on the saved content, not the rendered page: running the_content would add
 * whatever other plugins append (share buttons, ads, related posts) and is slow on
 * large archives. Block comments are dropped, shortcode tags are dropped but the text
 * they wrap is kept (page builders put the content inside them), and headings become
 * "## " markers so the chunker can split on sections.
 */
final class Extractor
{
    /**
     * @return array{title: string, body: string, metadata: array<string, mixed>}
     */
    public static function extract(\WP_Post $post): array
    {
        $fields = Settings::fields();
        $title  = in_array('title', $fields, true) || $post->post_type === 'attachment'
            ? self::clean(get_the_title($post))
            : '';

        $lead = [];
        if ($post->post_type === 'attachment') {
            $alt = self::clean((string) get_post_meta($post->ID, '_wp_attachment_image_alt', true));
            if ($alt !== '') {
                $lead[] = 'Image: ' . $alt;
            }
            if (trim($post->post_excerpt) !== '') {
                $lead[] = 'Caption: ' . self::clean($post->post_excerpt);
            }
        } else {
            if (in_array('excerpt', $fields, true) && trim($post->post_excerpt) !== '') {
                $lead[] = 'Summary: ' . self::clean($post->post_excerpt);
            }
            if (in_array('meta_description', $fields, true)) {
                $key  = Backfill::metaDescriptionKey();
                $desc = $key !== '' ? self::clean((string) get_post_meta($post->ID, $key, true)) : '';
                if ($desc !== '') {
                    $lead[] = 'Description: ' . $desc;
                }
            }
        }
        if (in_array('terms', $fields, true)) {
            $names = self::termNames($post);
            if ($names !== []) {
                $lead[] = 'Topics: ' . implode(', ', $names);
            }
        }
        if (in_array('author', $fields, true) && $post->post_type !== 'attachment') {
            $author = self::clean((string) get_the_author_meta('display_name', (int) $post->post_author));
            if ($author !== '') {
                $lead[] = 'Author: ' . $author;
            }
        }
        if (in_array('custom_fields', $fields, true)) {
            foreach (Settings::metaKeys() as $key) {
                $value = self::metaText(get_post_meta($post->ID, $key, false));
                if ($value !== '') {
                    $lead[] = self::label($key) . ': ' . $value;
                }
            }
        }

        $content = in_array('content', $fields, true) || $post->post_type === 'attachment'
            ? self::contentToText((string) $post->post_content)
            : '';
        $body = trim(implode("\n", $lead) . ($lead !== [] && $content !== '' ? "\n\n" : '') . $content);

        /**
         * Filters the text embedded for a post, before chunking. Headings are "## " lines.
         *
         * @param string   $body The text.
         * @param \WP_Post $post The post or attachment.
         */
        $body = (string) apply_filters('aipcf_embed_text', $body, $post);

        return ['title' => $title, 'body' => $body, 'metadata' => self::metadata($post)];
    }

    /**
     * Words in a text: runs of letters and digits, with the marks that belong to them, so
     * Marathi and Hindi (whose vowel signs are marks) count like English. Scripts written
     * without spaces (Chinese, Japanese, Thai) count each run as one word.
     */
    public static function words(string $text): int
    {
        return (int) preg_match_all('/[\p{L}\p{N}][\p{L}\p{N}\p{M}\'\x{2019}-]*/u', $text);
    }

    /**
     * Saved post content to plain text with "## " headings and blank-line paragraphs.
     */
    public static function contentToText(string $html): string
    {
        $html = (string) preg_replace('/<!--.*?-->/s', '', $html);                     // block delimiters, comments
        $html = (string) preg_replace('#<(script|style|noscript|svg|iframe)\b.*?</\1>#is', '', $html);
        $html = (string) preg_replace('/\[\/?[A-Za-z][\w-]*(?:\s[^\]]*)?\]/', ' ', $html);   // shortcode tags, keep inner text
        $html = (string) preg_replace_callback('#<table\b.*?</table>#is', [self::class, 'tableToText'], $html);
        $html = (string) preg_replace_callback(
            '#<h[1-6]\b[^>]*>(.*?)</h[1-6]>#is',
            static fn(array $m): string => "\n\n## " . trim(self::clean($m[1])) . "\n\n",
            $html
        );
        $html = (string) preg_replace('#<(br|hr)\b[^>]*>#i', "\n", $html);
        $html = (string) preg_replace('#</?(p|div|section|article|blockquote|figure|figcaption|ul|ol|pre|details|summary)\b[^>]*>#i', "\n\n", $html);
        $html = (string) preg_replace('#<li\b[^>]*>#i', "\n- ", $html);

        $text = html_entity_decode(wp_strip_all_tags($html, false), ENT_QUOTES | ENT_HTML5, 'UTF-8');
        $text = str_replace("\xC2\xA0", ' ', $text);
        $text = (string) preg_replace('/[ \t]+/', ' ', $text);
        $text = (string) preg_replace('/ *\n */', "\n", $text);
        $text = (string) preg_replace('/\n{3,}/', "\n\n", $text);
        $text = (string) preg_replace("/^## *\n/m", '', $text);                           // empty headings
        return trim($text);
    }

    /**
     * A table as one line per row, cells separated by " | ", whatever whitespace the HTML has.
     *
     * @param array<int, string> $m
     */
    private static function tableToText(array $m): string
    {
        $rows = [];
        foreach (preg_split('#<tr\b[^>]*>#i', $m[0]) ?: [] as $row) {
            $cells = [];
            foreach (preg_split('#</t[dh]>#i', $row) ?: [] as $cell) {
                $cell = self::clean($cell);
                if ($cell !== '') {
                    $cells[] = $cell;
                }
            }
            if ($cells !== []) {
                $rows[] = implode(' | ', $cells);
            }
        }
        return "\n\n" . implode("\n", $rows) . "\n\n";
    }

    /**
     * Stored with every chunk of the post. Keys with metadata indexes in the store
     * (site, post_type, object_id, lang) are filterable.
     *
     * @return array<string, mixed>
     */
    public static function metadata(\WP_Post $post): array
    {
        $metadata = [
            'site'        => self::siteKey(),
            'object_type' => $post->post_type === 'attachment' ? 'attachment' : 'post',
            'object_id'   => (int) $post->ID,
            'post_type'   => $post->post_type,
            'title'       => mb_substr(self::clean(get_the_title($post)), 0, 200),
            'url'         => $post->post_type === 'attachment' ? (string) wp_get_attachment_url($post->ID) : (string) get_permalink($post),
            'date'        => (string) $post->post_date_gmt,
            // No modified time: it changes on every save and is part of the content
            // hash, so a save without edits would be embedded (and billed) again.
            'lang'        => self::lang($post),
        ];
        /**
         * Filters the metadata stored with each chunk of a post (text is added per chunk).
         *
         * @param array    $metadata
         * @param \WP_Post $post
         */
        return (array) apply_filters('aipcf_embed_metadata', $metadata, $post);
    }

    /**
     * Short, stable ID for this site, so one index can hold several sites.
     */
    public static function siteKey(): string
    {
        return substr(md5(home_url() . '|' . get_current_blog_id()), 0, 8);
    }

    private static function lang(\WP_Post $post): string
    {
        $locale = (string) apply_filters('aipcf_embed_locale', get_locale(), $post);
        return strtolower((string) strtok($locale, '_-'));
    }

    /**
     * @return list<string>
     */
    private static function termNames(\WP_Post $post): array
    {
        $names = [];
        foreach (get_object_taxonomies($post->post_type, 'objects') as $taxonomy) {
            if (!$taxonomy->public) {
                continue;
            }
            $terms = get_the_terms($post, $taxonomy->name);
            if (is_array($terms)) {
                foreach ($terms as $term) {
                    if ($term->slug !== 'uncategorized') {
                        $names[] = self::clean($term->name);
                    }
                }
            }
        }
        return array_values(array_unique(array_filter($names)));
    }

    /**
     * @param array<int, mixed> $values
     */
    private static function metaText(array $values): string
    {
        $out = [];
        array_walk_recursive($values, static function ($v) use (&$out): void {
            if (is_scalar($v) && !is_bool($v)) {
                $t = self::clean((string) $v);
                // Skip empty values and long numbers (IDs, timestamps), which carry no meaning.
                if ($t !== '' && !(is_numeric($t) && strlen($t) >= 10)) {
                    $out[] = $t;
                }
            }
        });
        return mb_substr(implode(', ', array_unique($out)), 0, 2000);
    }

    private static function label(string $metaKey): string
    {
        return ucfirst(trim(str_replace(['_', '-'], ' ', $metaKey)));
    }

    private static function clean(string $text): string
    {
        $text = html_entity_decode(wp_strip_all_tags($text), ENT_QUOTES | ENT_HTML5, 'UTF-8');
        return trim((string) preg_replace('/\s+/u', ' ', str_replace("\xC2\xA0", ' ', $text)));
    }
}