Optimasi Performa AI Agent PHP

Terakhir diperbarui: 22 Agustus 2026

Diagram Arsitektur: Pilar Optimasi Performa & Reduksi Latensi AI Agent PHP

WebP Lossless • 80 KB • Schema Ready

Teknik Response Caching pada AI Agent PHP

Direct Answer:Response Caching menyimpan hasil inferensi model berdasarkan hash deterministik dari task prompt dan skema tools yang diizinkan, sehingga request berulang dapat disajikan secara instan tanpa memanggil API OpenAI kembali.

Berikut adalah implementasi class RedisResponseCache berkecepatan tinggi dengan auto-invalidation TTL:

<?php

namespace App\Optimization;

use Redis;

/**
 * Class RedisResponseCache
 * Mengelola in-memory caching hasil inferensi LLM untuk memangkas latensi prompt berulang.
 */
class RedisResponseCache
{
    private Redis $redis;
    private int $defaultTtl;

    public function __construct(Redis $redis, int $ttlSeconds = 3600)
    {
        $this->redis      = $redis;
        $this->defaultTtl = $ttlSeconds;
    }

    /**
     * Menghasilkan hash SHA-256 identitas unik dari prompt dan konfigurasi tools.
     */
    private function generateKey(string $task, array $tools = []): string
    {
        // Normalisasi teks prompt menjadi lowercase untuk memaksimalkan cache hit ratio
        $payload = strtolower(trim($task)) . '|' . json_encode($tools);
        return 'agent_cache:' . hash('sha256', $payload);
    }

    /**
     * Mengambil respons cache dari Redis (Waktu akses ~12ms).
     */
    public function get(string $task, array $tools = []): ?string
    {
        $key    = $this->generateKey($task, $tools);
        $cached = $this->redis->get($key);

        return $cached !== false ? (string)$cached : null;
    }

    /**
     * Menyimpan respons agen cerdas ke dalam Redis dengan masa kedaluwarsa (TTL).
     */
    public function set(string $task, string $response, array $tools = [], ?int $ttl = null): void
    {
        $key    = $this->generateKey($task, $tools);
        $expiry = $ttl ?? $this->defaultTtl;
        $this->redis->setex($key, $expiry, $response);
    }
}

// ==========================================
// Integrasi ke dalam loop Orchestrator Agen:
// ==========================================
$cache  = new RedisResponseCache($redisClient, ttlSeconds: 1800);
$cached = $cache->get($userPrompt, $availableTools);

if ($cached !== null) {
    // Cache Hit: Segera kembalikan respons tanpa menghubungi API OpenAI
    return $cached . ' [Disajikan dari Redis Cache • 12ms]';
}

// Cache Miss: Jalankan inferensi LLM penuh
$freshResult = $agent->run($userPrompt);
$cache->set($userPrompt, $freshResult, $availableTools);
return $freshResult;

Streaming Response Real-time dengan Server-Sent Events (SSE)

Direct Answer:Server-Sent Events (SSE) mentransmisikan token jawaban secara sekuensial (*chunk-by-chunk*) langsung ke browser segera setelah model LLM menghasilkan kata pertama (TTFT < 200ms) tanpa menunggu seluruh paragraf selesai.

Kode streaming endpoint PHP dengan header anti-buffering Nginx:

<?php
// =========================================================================
// stream.php - Server-Sent Events (SSE) Endpoint untuk AI Agent Response
// =========================================================================

// 1. Header Wajib untuk Streaming HTTP SSE
header('Content-Type: text/event-stream');
header('Cache-Control: no-cache, no-transform');
header('Connection: keep-alive');
header('X-Accel-Buffering: no'); // Wajib: Matikan buffering proxy Nginx

// 2. Nonaktifkan semua layer Output Buffering bawaan PHP
while (ob_get_level() > 0) {
    ob_end_clean();
}

$payload = json_encode([
    'model'    => 'gpt-4o-mini',
    'messages' => [
        ['role' => 'system', 'content' => 'Anda adalah asisten AI responsif.'],
        ['role' => 'user', 'content' => $_POST['prompt'] ?? 'Jelaskan komputasi awan']
    ],
    'stream'   => true, // Mengaktifkan streaming chunked transfer dari OpenAI
]);

$ch = curl_init('https://api.openai.com/v1/chat/completions');
curl_setopt_array($ch, [
    CURLOPT_POST           => true,
    CURLOPT_POSTFIELDS     => $payload,
    CURLOPT_HTTPHEADER     => [
        'Authorization: Bearer ' . getenv('OPENAI_API_KEY'),
        'Content-Type: application/json',
    ],
    CURLOPT_RETURNTRANSFER => false, // Callback write langsung menangani output
    CURLOPT_WRITEFUNCTION  => function ($ch, $chunk) {
        $lines = explode("\n", $chunk);
        foreach ($lines as $line) {
            $line = trim($line);
            // Tangkap baris data transmisi SSE
            if (str_starts_with($line, 'data: ') && $line !== 'data: [DONE]') {
                $json  = json_decode(substr($line, 6), true);
                $delta = $json['choices'][0]['delta']['content'] ?? '';
                if ($delta !== '') {
                    // Dorong token chunk langsung ke socket browser klien
                    echo "data: " . json_encode(['text' => $delta]) . "\n\n";
                    flush();
                }
            }
        }
        return strlen($chunk);
    },
]);

curl_exec($ch);
curl_close($ch);

// Sinyal penutup transmisi streaming SSE
echo "data: [DONE]\n\n";
flush();

Prompt Token Pruning dan Kompresi Konteks AI Agent

Direct Answer:Prompt Token Pruning adalah teknik optimasi komputasi yang memangkas teks berlebih, membuang spasi ganda, dan menerapkan Sliding Window Context Truncation pada riwayat percakapan agar ukuran payload token ke model LLM tetap ramping tanpa mengurangi akurasi penalaran agen.

Berikut adalah implementasi class PromptOptimizer untuk memangkas token prompt dan riwayat percakapan:

<?php

namespace App\Optimization;

class PromptOptimizer
{
    /**
     * Memangkas whitespace berlebih, markdown redundant, dan komentar kosong.
     */
    public static function pruneText(string $rawText): string
    {
        // 1. Konversi multiple newline menjadi maksimal 2 newline
        $clean = preg_replace("/\n{3,}/", "\n\n", $rawText);
        // 2. Hapus spasi horizontal berlebih di setiap baris
        $clean = preg_replace("/[ \t]{2,}/", " ", $clean);
        return trim($clean);
    }

    /**
     * Menerapkan Sliding Window Context Truncation pada riwayat pesan (Chat History).
     * Mempertahankan System Prompt (index 0) dan N turn pesan interaksi terbaru.
     */
    public static function truncateHistory(array $messages, int $maxHistoryTurns = 6): array
    {
        if (count($messages) <= $maxHistoryTurns + 1) {
            return $messages;
        }

        $systemPrompt = $messages[0]; // Selalu pertahankan System Prompt dasar
        $recentTurns  = array_slice($messages, -$maxHistoryTurns);

        return array_merge([$systemPrompt], $recentTurns);
    }

    /**
     * Semantic Chunking: Memotong teks dokumen panjang berdasarkan batas semantik paragraf/kalimat.
     */
    public static function chunkDocument(string $document, int $maxCharsPerChunk = 1500): array
    {
        $paragraphs = explode("\n\n", self::pruneText($document));
        $chunks     = [];
        $current    = "";

        foreach ($paragraphs as $para) {
            if (mb_strlen($current . $para) > $maxCharsPerChunk && !empty($current)) {
                $chunks[] = trim($current);
                $current  = $para . "\n\n";
            } else {
                $current .= $para . "\n\n";
            }
        }

        if (!empty(trim($current))) {
            $chunks[] = trim($current);
        }

        return $chunks;
    }
}

Implementasi Async Queue Redis untuk AI Agent PHP

Direct Answer:Antrean asinkron (Async Queue) memisahkan penerimaan request HTTP dari eksekusi tugas agen berat, membebaskan web server untuk segera mengembalikan Job ID ke klien dalam hitungan milidetik.

Pengiriman job via Redis RPUSH dan pemrosesan worker non-blocking via BLPOP:

<?php

// =========================================================================
// 1. Dispatcher Controller (api/dispatch_task.php) - Respon Instan < 50ms
// =========================================================================
$redis = new Redis();
$redis->connect('127.0.0.1', 6379);

$jobId   = 'job_' . bin2hex(random_bytes(16));
$payload = json_encode([
    'id'         => $jobId,
    'task'       => $_POST['task'] ?? '',
    'user_id'    => $authenticatedUserId,
    'created_at' => time(),
]);

// Masukkan payload job ke antrean antrean Redis
$redis->rPush('agent:tasks_queue', $payload);

echo json_encode([
    'status'   => 'queued',
    'job_id'   => $jobId,
    'poll_url' => "/api/check_status.php?job_id={$jobId}",
]);

// =========================================================================
// 2. Background Worker Daemon (worker.php) - Dikelola oleh Supervisor
// =========================================================================
while (true) {
    // Tunggu antrean task baru dengan blocking pop (timeout 5 detik)
    $job = $redis->blPop(['agent:tasks_queue'], 5);
    if (!$job) {
        // Bebaskan cycle memory pada perulangan idle
        gc_collect_cycles();
        continue;
    }

    $data   = json_decode($job[1], true);
    $jobId  = $data['id'];
    $task   = $data['task'];

    try {
        // Eksekusi tugas inferensi AI berat di background
        $result = $agent->run($task);
        // Simpan hasil ke Redis dengan masa kedaluwarsa 2 jam
        $redis->setex("agent:result:{$jobId}", 7200, json_encode(['status' => 'completed', 'result' => $result]));
    } catch (Throwable $e) {
        $redis->setex("agent:result:{$jobId}", 7200, json_encode(['status' => 'failed', 'error' => $e->getMessage()]));
    } finally {
        // Garansi pembersihan alokasi RAM per eksekusi task
        gc_collect_cycles();
    }
}

Implementasi Connection Pooling: Swoole PDOPool & PDO Persistence

Direct Answer:Connection Pooling mendaur ulang koneksi database TCP yang sudah terbuka antar request menggunakan coroutine pool Swoole\Database\PDOPool pada server asinkron atau atribut PDO::ATTR_PERSISTENT => true pada runtime PHP-FPM, menghemat 30-50ms overhead handshake per transaksi agen.

Berikut adalah contoh kode implementasi spesifik menggunakan Swoole\Database\PDOPool pada HTTP Server Coroutine dan fallback ke PDO Persistent Connection:

<?php

// =========================================================================
// Opsi A: High-Concurrency Coroutine Server dengan Swoole\Database\PDOPool
// =========================================================================

use Swoole\Http\Server;
use Swoole\Http\Request;
use Swoole\Http\Response;
use Swoole\Database\PDOConfig;
use Swoole\Database\PDOPool;
use Swoole\Coroutine;

// 1. Inisialisasi Konfigurasi Driver Database PDO Swoole
$pdoConfig = (new PDOConfig())
    ->withHost(getenv('DB_HOST') ?: '127.0.0.1')
    ->withPort(3306)
    ->withDbName(getenv('DB_NAME') ?: 'ai_agent_db')
    ->withCharset('utf8mb4')
    ->withUsername(getenv('DB_USER') ?: 'root')
    ->withPassword(getenv('DB_PASS') ?: '')
    ->withOptions([
        PDO::ATTR_ERRMODE            => PDO::ERRMODE_EXCEPTION,
        PDO::ATTR_DEFAULT_FETCH_MODE => PDO::FETCH_ASSOC,
        PDO::ATTR_TIMEOUT            => 5,
    ]);

// 2. Buat Connection Pool dengan kapasitas maksimal 64 koneksi aktif
$pool = new PDOPool($pdoConfig, 64);

$server = new Server('0.0.0.0', 9501);

$server->on('Request', function (Request $request, Response $response) use ($pool) {
    // Tangkap koneksi dari pool secara non-blocking
    $pdo = $pool->get();

    try {
        // Query riwayat context session agen dalam hitungan mikrodetik
        $stmt = $pdo->prepare('SELECT * FROM agent_sessions WHERE session_id = ? LIMIT 1');
        $stmt->execute([$request->get['session_id'] ?? 'guest']);
        $sessionData = $stmt->fetch();

        $response->header('Content-Type', 'application/json');
        $response->end(json_encode([
            'status'  => 'success',
            'latency' => '<1ms (Connection Pool Reused)',
            'session' => $sessionData,
        ]));
    } catch (Throwable $e) {
        $response->status(500);
        $response->end(json_encode(['error' => $e->getMessage()]));
    } finally {
        // Wajib: Kembalikan koneksi PDO ke dalam pool untuk request coroutine lain
        $pool->put($pdo);
    }
});

// $server->start(); // Jalankan server Swoole

<?php

namespace App\Database;

use PDO;
use PDOException;

/**
 * Opsi B: Persistent Connection Manager untuk Lingkungan Standar PHP-FPM / Nginx
 */
class PersistentConnectionPool
{
    private static ?PDO $instance = null;

    /**
     * Mengembalikan instance PDO Persistent Connection yang di-reuse lintas request worker.
     */
    public static function getConnection(): PDO
    {
        if (self::$instance === null) {
            $dsn = sprintf(
                'mysql:host=%s;dbname=%s;charset=utf8mb4',
                getenv('DB_HOST') ?: '127.0.0.1',
                getenv('DB_NAME') ?: 'ai_agent_db'
            );

            $options = [
                // Mengaktifkan koneksi persisten lintas lifecycle FPM worker
                PDO::ATTR_PERSISTENT         => true,
                PDO::ATTR_ERRMODE            => PDO::ERRMODE_EXCEPTION,
                PDO::ATTR_DEFAULT_FETCH_MODE => PDO::FETCH_ASSOC,
                PDO::ATTR_EMULATE_PREPARES   => false,
                PDO::ATTR_TIMEOUT            => 5,
            ];

            try {
                self::$instance = new PDO(
                    $dsn,
                    getenv('DB_USER') ?: 'root',
                    getenv('DB_PASS') ?: '',
                    $options
                );
            } catch (PDOException $e) {
                throw new \RuntimeException("Database Connection Failed: " . $e->getMessage());
            }
        }

        return self::$instance;
    }
}

FAQ: Optimasi Performa & Latensi AI Agent PHP