ChatbotAI/app/Services/KnowledgeBase/RAGService.php

<?php

namespace App\Services\KnowledgeBase;

use App\Services\Ollama\OllamaService;
use App\Services\Qdrant\QdrantService;
use Illuminate\Support\Facades\Log;
use RuntimeException;

/**
 * RAGService (Retrieval-Augmented Generation)
 *
 * Koordinasi proses RAG:
 * 1. Jana embedding untuk soalan user
 * 2. Cari context paling relevan dari Qdrant
 * 3. Bina context string
 * 4. Hantar ke Ollama untuk jawapan
 * 5. Return jawapan + source references
 */
class RAGService
{
    private int $maxContextChunks;
    private int $maxContextWords;

    public function __construct(
        private readonly OllamaService $ollama,
        private readonly QdrantService $qdrant,
    ) {
        $this->maxContextChunks = config('knowledgebase.rag.max_context_chunks', 5);
        $this->maxContextWords  = config('knowledgebase.rag.max_context_words', 2000);
    }

    /**
     * Jawab soalan menggunakan RAG.
     *
     * @param  string $question      Soalan pengguna
     * @param  ?int   $categoryId    Filter kategori (null = semua)
     * @return array{
     *   answer: string,
     *   has_answer: bool,
     *   sources: array[],
     *   context_chunks: array[],
     *   model_used: string,
     *   tokens_used: ?int,
     *   response_time: float
     * }
     * @throws RuntimeException Jika Ollama atau Qdrant tidak tersedia
     */
    public function ask(string $question, ?int $categoryId = null): array
    {
        $startTime = microtime(true);

        // ── Langkah 1: Jana embedding untuk soalan ─────────────────────────
        $queryVector = $this->ollama->embed($question);

        // ── Langkah 2: Cari context relevan dari Qdrant ─────────────────────
        $filter = $this->qdrant->buildFilter(
            categoryId: $categoryId,
            isActive: true,
        );

        $scoreThreshold = config('qdrant.search.score_threshold', 0.3);

        $searchResults = $this->qdrant->searchSimilar(
            vector: $queryVector,
            limit: $this->maxContextChunks,
            filter: $filter,
            scoreThreshold: $scoreThreshold,
        );

        //log search result
        \Log::info('Qdrant search raw results', [
            'question' => $question,
            'results' => $searchResults,
        ]);

        \Log::info('Qdrant raw results', [
            'scores' => array_map(fn($r) => $r['score'] ?? null, $searchResults),
        ]);

        if (empty($searchResults)) {
            $responseTime = round(microtime(true) - $startTime, 3);

            return [
                'answer'         => config('ollama.rag_system_prompt_no_result',
                    'Maaf, saya tidak menemui maklumat berkaitan dalam pangkalan pengetahuan kami. ' .
                    'Sila hubungi pejabat kami untuk maklumat lanjut.'),
                'has_answer'     => false,
                'sources'        => [],
                'context_chunks' => [],
                'model_used'     => config('ollama.chat_model'),
                'tokens_used'    => null,
                'response_time'  => $responseTime,
            ];
        }

        // ── Langkah 3: Bina context string ─────────────────────────────────
        [$context, $contextChunksData] = $this->buildContext($searchResults);

        // ── Langkah 4: Hantar ke Ollama ─────────────────────────────────────
        $chatResult = $this->ollama->chat($question, $context);

        // ── Langkah 5: Bina source references ──────────────────────────────
        $sources = $this->buildSourceReferences($searchResults);

        $responseTime = round(microtime(true) - $startTime, 3);

        // Tentukan sama ada model ada jawapan atau tidak
        $hasAnswer = $this->detectHasAnswer($chatResult['answer']);

        return [
            'answer'         => $chatResult['answer'],
            'has_answer'     => $hasAnswer,
            'sources'        => $sources,
            'context_chunks' => $contextChunksData,
            'model_used'     => $chatResult['model'],
            'tokens_used'    => $chatResult['tokens'],
            'response_time'  => $responseTime,
        ];
    }

    /**
     * Bina context string dari search results.
     * Had bilangan perkataan supaya tidak melebihi context window model.
     *
     * @return array{0: string, 1: array[]}
     */
    private function buildContext(array $searchResults): array
    {
        $contextParts   = [];
        $chunksData     = [];
        $totalWords     = 0;

        foreach ($searchResults as $result) {
            $payload = $result['payload'] ?? [];
            $text    = $payload['text'] ?? '';

            if (empty($text)) {
                continue;
            }

            $words = str_word_count($text);

            if ($totalWords + $words > $this->maxContextWords) {
                // Potong jika context dah terlalu panjang
                if (empty($contextParts)) {
                    // Sekurang-kurangnya masukkan satu chunk
                    $contextParts[] = $text;
                    $chunksData[]   = $this->extractChunkData($result);
                }
                break;
            }

            $source = $this->formatSourceLabel($payload);
            $contextParts[] = "[Sumber: {$source}]\n{$text}";
            $chunksData[]   = $this->extractChunkData($result);
            $totalWords    += $words;
        }

        return [implode("\n\n---\n\n", $contextParts), $chunksData];
    }

    /**
     * Bina array source references untuk paparan kepada pengguna.
     */
    private function buildSourceReferences(array $searchResults): array
    {
        $sources = [];
        $seen    = []; // Elak duplikasi sumber yang sama

        foreach ($searchResults as $result) {
            $payload = $result['payload'] ?? [];

            $sourceKey = ($payload['document_id'] ?? '') . '_' .
                         ($payload['knowledge_item_id'] ?? '') . '_' .
                         ($payload['page_number'] ?? '');

            if (isset($seen[$sourceKey])) {
                continue;
            }

            $seen[$sourceKey] = true;

            $sources[] = [
                'type'               => $payload['source_type'] ?? 'unknown',
                'knowledge_type'     => $payload['knowledge_type'] ?? '',
                'title'              => $payload['title'] ?? 'Tiada tajuk',
                'category'           => $payload['category_name'] ?? '',
                'category_id'        => $payload['category_id'] ?? null,
                'page_number'        => $payload['page_number'] ?? null,
                'section_heading'    => $payload['section_heading'] ?? null,
                'document_id'        => $payload['document_id'] ?? null,
                'knowledge_item_id'  => $payload['knowledge_item_id'] ?? null,
                'score'              => round($result['score'] ?? 0, 4),
            ];
        }

        return $sources;
    }

    /**
     * Extract data chunk untuk disimpan dalam chat_logs.
     */
    private function extractChunkData(array $result): array
    {
        return [
            'point_id'    => $result['id'] ?? null,
            'score'       => round($result['score'] ?? 0, 4),
            'title'       => $result['payload']['title'] ?? '',
            'category'    => $result['payload']['category_name'] ?? '',
            'source_type' => $result['payload']['source_type'] ?? '',
            'page_number' => $result['payload']['page_number'] ?? null,
        ];
    }

    private function formatSourceLabel(array $payload): string
    {
        $title    = $payload['title'] ?? 'Tanpa tajuk';
        $page     = isset($payload['page_number']) ? ", ms. {$payload['page_number']}" : '';
        $category = $payload['category_name'] ?? '';

        return "{$title}{$page} ({$category})";
    }

    /**
     * Detect sama ada model sebenarnya ada jawapan atau tidak.
     * Semak jika jawapan adalah "tidak tahu" / fallback.
     */
    private function detectHasAnswer(string $answer): bool
    {
        $noAnswerPatterns = [
            'tidak menemui',
            'tiada maklumat',
            'tidak terdapat dalam',
            'sila hubungi',
            'tidak dapat menjawab',
            'maklumat tidak tersedia',
        ];

        $answerLower = mb_strtolower($answer);
        foreach ($noAnswerPatterns as $pattern) {
            if (str_contains($answerLower, $pattern)) {
                return false;
            }
        }

        return !empty(trim($answer));
    }
}