<?php
declare(strict_types=1);

/** Deterministic candidate retrieval, not semantic answer or citation prediction. */
final class TheDotDevAgenticBridge_Question_Evidence
{
    public static function terms(string $question): array
    {
        $words = preg_split('/[^\p{L}\p{N}]+/u', mb_strtolower($question)) ?: [];
        $stop = ['what','when','where','which','does','with','from','your','have','that','this','they','their','about','could','would','should','there','offer','please','can','you','the','and','are','for','how'];
        return array_slice(array_values(array_unique(array_filter($words, static fn($w) => mb_strlen($w) >= 3 && !in_array($w, $stop, true)))), 0, 20);
    }

    public static function candidates(string $question, array $pages): array
    {
        $terms = self::terms($question);
        $results = [];
        foreach (array_slice($pages, 0, 50) as $page) {
            $text = mb_substr((string) $page['text'], 0, 20000);
            $haystack = mb_strtolower($text);
            $matches = array_values(array_filter($terms, static fn($term) => preg_match('/(?<![\p{L}\p{N}])' . preg_quote($term, '/') . '(?![\p{L}\p{N}])/u', $haystack) === 1));
            if ($matches === []) continue;
            // Use the same word boundaries for the quote as for candidate matching.
            preg_match('/(?<![\p{L}\p{N}])' . preg_quote($matches[0], '/') . '(?![\p{L}\p{N}])/iu', $text, $quote_match, PREG_OFFSET_CAPTURE);
            $position = mb_strlen(substr($text, 0, $quote_match[0][1]));
            $page['excerpt'] = mb_substr($text, max(0, (int) $position - 100), 500);
            unset($page['text']);
            $page['matches'] = $matches;
            $page['unmatched'] = array_values(array_diff($terms, $matches));
            $results[] = $page;
        }
        usort($results, static fn($a, $b) => count($b['matches']) <=> count($a['matches']) ?: $a['id'] <=> $b['id']);
        return array_slice($results, 0, 5);
    }
}