<?php
// SPDX-License-Identifier: GPL-3.0-or-later
declare(strict_types=1);
namespace Aiwend\Chat;

use Aiwend\Db\Connection;
use Aiwend\Policy\Capabilities;
use Aiwend\Policy\Principal;
use Aiwend\Support\Clock;

/**
 * Stage D read model for the administrator: how many answers found no source, how visitors rated answers, which unanswered
 * questions recur. Nothing is stored: every figure is computed on request from the conversation tables, bounded by a time
 * window and a row cap, so it can never outlive the conversations it was computed from (retention, visitor deletion and
 * WordPress erasure apply unchanged). Question text is returned only to a user who may read conversations.
 *
 * Nothing here says a customer's problem was resolved: the product has no evidence of that. "No source" means retrieval
 * found no qualifying Knowledge for the question (the assistant may still be right to be silent, or retrieval may have missed).
 */
final class SupportInsights
{
    /** Minimum occurrences for a recurring group. This is a noise filter, not anonymization. */
    public const GROUP_MIN = 3;
    public const SAMPLE_CAP = 500;
    public const GROUP_LIMIT = 20;
    public const LABEL_CHARS = 120;
    public const SOURCE_CONVERSATIONS = 3;

    /** SQL (alias m, an assistant row): exact structured outcome, or the older structural reading described below. */
    public const SQL_EXACT = "m.outcome IN ('no_match', 'withdrawn')";
    /**
     * Rows stored before the structured outcome: a grounded assistant that sent no source to the model recorded an empty
     * provenance list (messages.grounding = '', v11). That is structural, not language-dependent, but it is an inference:
     * rows from before v11 have no provenance and are not recognised, and the inference is always labelled as such.
     */
    public const SQL_INFERRED = "(m.outcome IS NULL OR m.outcome = 'completed') AND m.grounding = ''";

    public function __construct(private Connection $db, private Clock $clock) {}

    private function authorize(Principal $principal): void
    {
        if (!$principal->isAuthenticated() || !$principal->can(Capabilities::EXPORT_CONVERSATIONS)) {
            throw new ChatException('forbidden', 403);
        }
    }

    /**
     * Counts since $since (inclusive). Definitions: conversations = sessions created in the period in which the visitor sent at least one message (a page view that only opened a session is not a conversation); answers = assistant
     * messages created in the period (a stopped answer counts as an answer); rateable = answers that are not stopped;
     * rated = answers with a helpful or not-helpful rating; failures = audit events (one per failed request, not per message).
     *
     * @return array<string, int|bool>
     */
    public function summary(Principal $principal, int $since, int $bot = 0, int $auditDays = 30): array
    {
        $this->authorize($principal);
        $site = $principal->siteId;
        $botArgs = $bot > 0 ? [$bot] : [];
        $conversations = (int) ($this->db->selectOne(
            "SELECT COUNT(*) AS n FROM {chat_sessions} cs WHERE cs.site_id = ? AND cs.created_at >= ?" . ($bot > 0 ? ' AND cs.bot_id = ?' : '')
            . " AND EXISTS (SELECT 1 FROM {messages} um WHERE um.site_id = cs.site_id AND um.session_id = cs.id AND um.role = 'user')",
            [$site, $since, ...$botArgs]
        )['n'] ?? 0);
        $a = $this->db->selectOne(
            "SELECT COUNT(*) AS answers,
                COALESCE(SUM(CASE WHEN " . self::SQL_EXACT . " THEN 1 ELSE 0 END), 0) AS exact,
                COALESCE(SUM(CASE WHEN m.outcome = 'withdrawn' THEN 1 ELSE 0 END), 0) AS withdrawn,
                COALESCE(SUM(CASE WHEN " . self::SQL_INFERRED . " THEN 1 ELSE 0 END), 0) AS inferred,
                COALESCE(SUM(CASE WHEN m.outcome = 'stopped' THEN 1 ELSE 0 END), 0) AS stopped
             FROM {messages} m WHERE m.site_id = ? AND m.role = 'assistant' AND m.created_at >= ?" . ($bot > 0 ? ' AND m.bot_id = ?' : ''),
            [$site, $since, ...$botArgs]
        ) ?? [];
        $rating = [1 => 0, -1 => 0];
        foreach ($this->db->select(
            'SELECT f.rating, COUNT(*) AS n FROM {messages} m JOIN {message_feedback} f ON f.site_id = m.site_id AND f.message_id = m.id
             WHERE m.site_id = ? AND m.created_at >= ?' . ($bot > 0 ? ' AND m.bot_id = ?' : '') . ' GROUP BY f.rating',
            [$site, $since, ...$botArgs]
        ) as $r) {
            $rating[(int) $r['rating']] = (int) $r['n'];
        }
        $answers = (int) ($a['answers'] ?? 0);
        $stopped = (int) ($a['stopped'] ?? 0);

        // Failures come from the activity log, which has its own retention and no assistant column: say how far back it reaches.
        $now = $this->clock->now();
        $failFrom = max($since, $now - max(1, $auditDays) * 86400);
        $failures = 0;
        $limits = 0;
        foreach ($this->db->select("SELECT outcome, COUNT(*) AS n FROM {audit} WHERE site_id = ? AND action = 'chat.send' AND created_at >= ? GROUP BY outcome", [$site, $failFrom]) as $r) {
            if (in_array($r['outcome'], ['failed_not_dispatched', 'failed_rejected', 'failed_usage_known', 'uncertain'], true)) {
                $failures += (int) $r['n'];
            } elseif ($r['outcome'] === 'budget_denied') {
                $limits += (int) $r['n'];
            }
        }
        return [
            'conversations' => $conversations,
            'answers' => $answers,
            'no_match_exact' => (int) ($a['exact'] ?? 0),
            'withdrawn' => (int) ($a['withdrawn'] ?? 0),
            'no_match_inferred' => (int) ($a['inferred'] ?? 0),
            'stopped' => $stopped,
            'rateable' => $answers - $stopped,
            'rated' => $rating[1] + $rating[-1],
            'positive' => $rating[1],
            'negative' => $rating[-1],
            'provider_failures' => $failures,
            'limit_refusals' => $limits,
            'failure_window_start' => $failFrom,
            'failure_partial' => $failFrom > $since,
        ];
    }

    /**
     * Recurring unanswered questions: the newest $cap answers that found no source, their questions normalized (lower case,
     * punctuation and spacing removed; e-mail addresses and long digit runs masked), exact matches after that counted.
     * Occurrences are questions asked, not people: one visitor can ask the same thing many times and a conversation can
     * contain several. Nothing is cached or stored.
     *
     * @return array{groups: list<array{label:string, occurrences:int, conversations:int, sessions:list<int>}>, analysed:int, capped:bool, minimum:int}
     */
    public function recurring(Principal $principal, int $since, int $bot = 0, int $cap = self::SAMPLE_CAP, int $minimum = self::GROUP_MIN): array
    {
        $this->authorize($principal);
        $cap = max(1, min(self::SAMPLE_CAP, $cap));
        $args = [$principal->siteId, $since];
        $botSql = '';
        if ($bot > 0) { $botSql = ' AND m.bot_id = ?'; $args[] = $bot; }
        $args[] = $cap + 1;
        $rows = $this->db->select(
            "SELECT u.content, m.session_id FROM {messages} m
             JOIN {messages} u ON u.site_id = m.site_id AND u.session_id = m.session_id AND u.role = 'user' AND u.request_id = m.request_id
             WHERE m.site_id = ? AND m.role = 'assistant' AND m.created_at >= ?$botSql AND ((" . self::SQL_EXACT . ') OR (' . self::SQL_INFERRED . ')) ORDER BY m.id DESC LIMIT ?',
            $args
        );
        $capped = count($rows) > $cap;
        $rows = array_slice($rows, 0, $cap);
        $groups = [];
        foreach ($rows as $r) {
            $label = self::normalize((string) $r['content']);
            if ($label === '') {
                continue;
            }
            $sid = (int) $r['session_id'];
            $groups[$label]['n'] = ($groups[$label]['n'] ?? 0) + 1;
            $groups[$label]['sessions'][$sid] = true;
        }
        $out = [];
        foreach ($groups as $label => $g) {
            if ($g['n'] >= $minimum) {
                $out[] = ['label' => (string) $label, 'occurrences' => $g['n'], 'conversations' => count($g['sessions']), 'sessions' => array_slice(array_keys($g['sessions']), 0, self::SOURCE_CONVERSATIONS)];
            }
        }
        usort($out, static fn(array $x, array $y): int => [$y['occurrences'], $x['label']] <=> [$x['occurrences'], $y['label']]);
        return ['groups' => array_slice($out, 0, self::GROUP_LIMIT), 'analysed' => count($rows), 'capped' => $capped, 'minimum' => $minimum];
    }

    /** Deterministic and local: no stemming, no translation, no similarity, so only questions that match after this step are grouped. */
    public static function normalize(string $text): string
    {
        $t = mb_strtolower(mb_substr($text, 0, 1000));
        $t = (string) preg_replace('/[^\s@]+@[^\s@]+\.[^\s@]+/u', ' email ', $t);
        $t = (string) preg_replace('/(?:\p{N}[\s().-]*){7,}/u', ' number ', $t);
        $t = (string) preg_replace('/[^\p{L}\p{N}]+/u', ' ', $t);
        return trim(mb_substr(trim($t), 0, self::LABEL_CHARS));
    }
}