{"id":2942,"date":"2026-11-11T09:41:12","date_gmt":"2026-11-11T02:41:12","guid":{"rendered":"https:\/\/komunitech.com\/blog\/?p=2942"},"modified":"2026-11-12T21:02:32","modified_gmt":"2026-11-12T14:02:32","slug":"mengatasi-error-parsing-json-dan-halusinasi-schema-pada-ai-agent","status":"publish","type":"post","link":"https:\/\/komunitech.com\/blog\/karir-skill\/mengatasi-error-parsing-json-dan-halusinasi-schema-pada-ai-agent\/","title":{"rendered":"Mengatasi Error Parsing JSON dan Halusinasi Schema pada AI Agent"},"content":{"rendered":"<p class=\"c9\">Membangun AI Agent yang andal untuk lingkungan produksi sering kali berbenturan dengan kenyataan pahit: Large Language Models (LLM) pada dasarnya adalah mesin pemprediksi token berturut-turut, bukan parser sintaksis yang deterministik. Saat kita meminta AI Agent memanggil alat (<span class=\"c6\">tool-calling<\/span><span class=\"c3\">) atau menghasilkan payload JSON untuk integrasi sistem, ketidakstabilan format sering menjadi batu sandungan utama.<\/span><\/p>\n<h2 id=\"h.btnp89buzgvy\" class=\"c19\"><span class=\"c10\">Apa itu Parsing Data?<\/span><\/h2>\n<p class=\"c9\"><span class=\"c3\">Parsing data adalah proses mengubah data mentah yang tidak terstruktur menjadi format terstruktur yang mudah dipahami dan diproses oleh komputer. <\/span><\/p>\n<p class=\"c9\"><span class=\"c3\">Proses parsing umumnya melalui beberapa tahapan utama:<\/span><\/p>\n<p class=\"c9\"><span class=\"c3\">1. Tokenisasi (Analisis Leksikal): Memecah teks atau kode menjadi bagian-bagian kecil (disebut token), seperti kata kunci, operator, atau nama variabel.<\/span><\/p>\n<p class=\"c9\"><span class=\"c3\">2. Analisis Sintaks: Memeriksa apakah urutan token tersebut sudah mengikuti aturan tata bahasa atau format yang benar.<\/span><\/p>\n<p class=\"c9\"><span class=\"c3\">3. Penyusunan Struktur (Parse Tree \/ AST): Membentuk representasi data yang terstruktur (seperti pohon sintaksis) agar bisa diproses lebih lanjut oleh program.<\/span><\/p>\n<p class=\"c9\">Lalu mengapa LLM sering mengalami kegagalan sintaksis, apa saja dampaknya terhadap arsitektur automasi, serta strategi konkret untuk memaksa model menghasilkan <span class=\"c6\">structured output<\/span><span class=\"c3\">\u00a0yang 100% patuh pada standar produksi.<\/span><\/p>\n<p class=\"related-article-link\">Artikel Terkait: <a href=\"https:\/\/komunitech.com\/blog\/karir-skill\/sertifikasi-prompt-engineering\/\">Sertifikasi Prompt Engineering: Mana yang Dilihat Rekruter (2026)<\/a><\/p>\n<h2 id=\"h.7cng5hidht13\" class=\"c19\"><span class=\"c10\">Akar Masalah: Mengapa Model Gagal Membentuk JSON Valid?<\/span><\/h2>\n<p class=\"c9\"><span class=\"c3\">Meskipun model AI generasi terbaru memiliki tingkat kecerdasan bernalar yang tinggi, format keluaran teksnya bersifat probabilistik. Karakteristik ini memicu beberapa kecenderungan bawaan yang merusak integritas payload JSON:<\/span><\/p>\n<ul class=\"c28 lst-kix_8j5rogrrfhvn-0 start\">\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Markdown Code Block Wrapper (<\/span><span class=\"c16 c29\">&#8220;`json<\/span><span class=\"c16\">)<\/span><br \/>\nSecara bawaan, banyak LLM dilatih untuk menyajikan jawaban yang mudah dibaca oleh manusia. Model cenderung membungkus objek JSON di dalam blok kode Markdown. Ketika parser eksternal menerima string tersebut secara mentah, fungsi <span class=\"c18\">JSON.parse()<\/span>\u00a0atau <span class=\"c18\">json.loads()<\/span><span class=\"c3\">\u00a0akan melempar exception karena karakter backtick bukan merupakan bagian dari spesifikasi RFC 8259.<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Teks Pengantar dan Penutup (<\/span><span class=\"c16 c6\">Conversational Filler<\/span><span class=\"c16\">)<\/span><br \/>\nModel sering kali menambahkan kalimat ramah seperti <span class=\"c6\">&#8220;Tentu, berikut adalah data JSON yang Anda minta:&#8221;<\/span>\u00a0di awal atau <span class=\"c6\">&#8220;Semoga membantu!&#8221;<\/span><span class=\"c3\">\u00a0di akhir pesan. Adanya teks bebas ini merusak struktur dokumen JSON secara instan.<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Penggunaan Trailing Comma<\/span><br \/>\nSesuai standar JSON formal, koma tidak boleh ditempatkan di akhir elemen terakhir dalam sebuah objek atau array (contoh: <span class=\"c18\">{&#8220;status&#8221;: &#8220;active&#8221;,}<\/span>). Namun, dalam banyak bahasa pemrograman (seperti JavaScript atau Python list), <span class=\"c6\">trailing comma<\/span><span class=\"c3\">\u00a0diizinkan. Model yang mempelajari miliaran baris kode sering kali menyisipkan koma ekstra ini tanpa disadari.<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Halusinasi Parameter dan Tipe Data pada Tool Definition<\/span><br \/>\nSaat melakukan <span class=\"c6\">function calling<\/span>, model dapat mengarang field yang sama sekali tidak dicantumkan dalam JSON Schema yang diberikan. Selain itu, model kerap mengalami tipe data yang tidak konsisten, seperti memberikan string <span class=\"c18\">&#8220;123&#8221;<\/span>\u00a0untuk field yang secara eksplisit meminta tipe integer <span class=\"c18\">123<\/span><span class=\"c3\">.<\/span><\/li>\n<\/ul>\n<h2 id=\"h.vjr9jju3x0di\" class=\"c19\"><span class=\"c10\">Dampak Schema Hallucination pada Workflow Automasi<\/span><\/h2>\n<p class=\"c9\">Ketidakpatuhan model terhadap format data bukan sekadar isu <span class=\"c6\">formatting<\/span><span class=\"c3\">, melainkan ancaman langsung terhadap keandalan sistem berskala besar.<\/span><\/p>\n<table class=\"c31\">\n<thead>\n<tr class=\"c26\">\n<td class=\"c17\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\"><span class=\"c2\">Jenis Dampak<\/span><\/p>\n<\/td>\n<td class=\"c17\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\"><span class=\"c2\">Mekanisme Kegagalan<\/span><\/p>\n<\/td>\n<td class=\"c17\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\"><span class=\"c2\">Konsekuensi Operasional<\/span><\/p>\n<\/td>\n<\/tr>\n<\/thead>\n<tbody>\n<tr class=\"c26\">\n<td class=\"c15\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\"><span class=\"c16\">API Failure (400 Bad Request)<\/span><\/p>\n<\/td>\n<td class=\"c15\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\"><span class=\"c3\">Payload yang dikirimkan AI Agent melanggar validasi schema pada REST API pihak ketiga.<\/span><\/p>\n<\/td>\n<td class=\"c15\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\">Request ditolak seketika oleh <span class=\"c6\">upstream server<\/span><span class=\"c3\">; transaksi bisnis gagal diproses.<\/span><\/p>\n<\/td>\n<\/tr>\n<tr class=\"c26\">\n<td class=\"c15\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\"><span class=\"c16\">Pipeline Terputus (Broken DAG)<\/span><\/p>\n<\/td>\n<td class=\"c15\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\">Exception yang tidak ditangani (<span class=\"c6\">unhandled parsing error<\/span><span class=\"c3\">) menghentikan eksekusi workflow pada node Agent.<\/span><\/p>\n<\/td>\n<td class=\"c15\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\"><span class=\"c3\">Sistem terhenti di tengah jalan; membutuhkan intervensi manual dari tim engineering.<\/span><\/p>\n<\/td>\n<\/tr>\n<tr class=\"c26\">\n<td class=\"c22\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\"><span class=\"c16\">Pemborosan Kredit API<\/span><\/p>\n<\/td>\n<td class=\"c22\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\">Agent terjebak dalam <span class=\"c6\">infinite loop<\/span>\u00a0saat mencoba mengulang permintaan yang gagal (<span class=\"c6\">retries<\/span><span class=\"c3\">).<\/span><\/p>\n<\/td>\n<td class=\"c22\" colspan=\"1\" rowspan=\"1\">\n<p class=\"c9\"><span class=\"c3\">Konsumsi token melonjak drastis; biaya operasional LLM membengkak tanpa hasil.<\/span><\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p class=\"related-article-link\">Artikel Terkait: <a href=\"https:\/\/komunitech.com\/blog\/karir-skill\/cara-menghasilkan-uang-dari-ai-2026\/\">Cara Menghasilkan Uang dari AI: Mana yang Beneran Ada Permintaannya (2026)<\/a><\/p>\n<h2 id=\"h.v33nkdq4q07t\" class=\"c19\"><span class=\"c10\">Solusi Ampuh Memaksa Output Terstruktur<\/span><\/h2>\n<p class=\"c9\">Untuk mengatasi masalah ini secara permanen, kita perlu menerapkan pendekatan berlapis (<span class=\"c6\">defense-in-depth<\/span><span class=\"c3\">) mulai dari tingkat inferensi model hingga validasi di sisi kode aplikasi.<\/span><\/p>\n<h3 id=\"h.uy2hpbl9xcx\" class=\"c14\"><span class=\"c24 c16\">1. Memanfaatkan Structured Outputs Bawaan (JSON Schema Mode)<\/span><\/h3>\n<p class=\"c9\">Mayoritas penyedia LLM utama kini menyediakan fitur <span class=\"c6\">Constrained Decoding<\/span><span class=\"c3\">\u00a0di tingkat API. Fitur ini memaksa pengolah dekoder LLM hanya memilih token yang valid berdasarkan grammar atau schema JSON yang didefinisikan.<\/span><\/p>\n<p class=\"c9\"><span class=\"c23\">\uec03<\/span><span class=\"c7 c27\"># Contoh implementasi JSON Schema Mode menggunakan OpenAI SDK<\/span><\/p>\n<blockquote>\n<p class=\"c9\"><span class=\"c13\">import <\/span><span class=\"c1\">json<\/span><\/p>\n<p class=\"c9\"><span class=\"c13\">from <\/span><span class=\"c4\">openai <\/span><span class=\"c13\">import <\/span><span class=\"c1\">OpenAI<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">client = OpenAI()<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">response = client.chat.completions.create(<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 model=<\/span><span class=\"c5\">&#8220;gpt-4o-mini&#8221;<\/span><span class=\"c1\">,<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 messages=[<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 {<\/span><span class=\"c5\">&#8220;role&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;system&#8221;<\/span><span class=\"c4\">, <\/span><span class=\"c5\">&#8220;content&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;Anda adalah asisten ekstraksi data.&#8221;<\/span><span class=\"c1\">},<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 {<\/span><span class=\"c5\">&#8220;role&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;user&#8221;<\/span><span class=\"c4\">, <\/span><span class=\"c5\">&#8220;content&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;Ekstrak data: Budi, 29 tahun, Software Engineer di Jakarta.&#8221;<\/span><span class=\"c1\">}<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 ],<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 response_format={<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;type&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;json_schema&#8221;<\/span><span class=\"c1\">,<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;json_schema&#8221;<\/span><span class=\"c1\">: {<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;name&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;user_profile&#8221;<\/span><span class=\"c1\">,<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;strict&#8221;<\/span><span class=\"c1\">: True,<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;schema&#8221;<\/span><span class=\"c1\">: {<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;type&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;object&#8221;<\/span><span class=\"c1\">,<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;properties&#8221;<\/span><span class=\"c1\">: {<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;nama&#8221;<\/span><span class=\"c4\">: {<\/span><span class=\"c5\">&#8220;type&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;string&#8221;<\/span><span class=\"c1\">},<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;umur&#8221;<\/span><span class=\"c4\">: {<\/span><span class=\"c5\">&#8220;type&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;integer&#8221;<\/span><span class=\"c1\">},<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;pekerjaan&#8221;<\/span><span class=\"c4\">: {<\/span><span class=\"c5\">&#8220;type&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;string&#8221;<\/span><span class=\"c1\">},<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;lokasi&#8221;<\/span><span class=\"c4\">: {<\/span><span class=\"c5\">&#8220;type&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;string&#8221;<\/span><span class=\"c1\">}<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 },<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;required&#8221;<\/span><span class=\"c4\">: [<\/span><span class=\"c5\">&#8220;nama&#8221;<\/span><span class=\"c4\">, <\/span><span class=\"c5\">&#8220;umur&#8221;<\/span><span class=\"c4\">, <\/span><span class=\"c5\">&#8220;pekerjaan&#8221;<\/span><span class=\"c4\">, <\/span><span class=\"c5\">&#8220;lokasi&#8221;<\/span><span class=\"c1\">],<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;additionalProperties&#8221;<\/span><span class=\"c1\">: False<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 }<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 \u00a0 \u00a0 }<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 }<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">)<\/span><\/p>\n<p class=\"c8\">\n<p class=\"c9\"><span class=\"c4\">data = json.loads(response.choices[<\/span><span class=\"c20\">0<\/span><span class=\"c4\">].message.content)<\/span><\/p>\n<p class=\"c9\"><span class=\"c23\">\uec02<\/span><\/p>\n<\/blockquote>\n<p class=\"c9\">Dengan mengaktifkan opsi <span class=\"c18\">&#8220;strict&#8221;: True<\/span><span class=\"c3\">, API jamin tidak akan menyisipkan markdown, teks pengantar, atau field yang tidak didaftarkan.<\/span><\/p>\n<h3 id=\"h.cpfsfbtahhh7\" class=\"c14\"><span class=\"c16 c24\">2. Pydantic Validation &amp; Instructor Pattern<\/span><\/h3>\n<p class=\"c9\">Di ekosistem Python, gabungan library Pydantic dan Instructor merupakan standar industri untuk mengelola <span class=\"c6\">structured outputs<\/span>. Pydantic menangani pengetikan tipe data (<span class=\"c6\">data typing<\/span><span class=\"c3\">) yang ketat dan otomatis melakukan konversi tipe jika memungkinkan.<\/span><\/p>\n<p class=\"c8\">\n<p class=\"c9\"><span class=\"c23\">\uec03<\/span><span class=\"c13\">import <\/span><span class=\"c1\">instructor<\/span><\/p>\n<p class=\"c9\"><span class=\"c13\">from <\/span><span class=\"c4\">pydantic <\/span><span class=\"c13\">import <\/span><span class=\"c1\">BaseModel, Field<\/span><\/p>\n<p class=\"c9\"><span class=\"c13\">from <\/span><span class=\"c4\">openai <\/span><span class=\"c13\">import <\/span><span class=\"c1\">OpenAI<\/span><\/p>\n<p class=\"c8\">\n<p class=\"c9\"><span class=\"c7 c27\"># Patch client OpenAI menggunakan Instructor<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">client = instructor.from_openai(OpenAI())<\/span><\/p>\n<p class=\"c8\">\n<p class=\"c9\"><span class=\"c13\">class <\/span><span class=\"c1\">UserProfile(BaseModel):<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 nama: <\/span><span class=\"c13\">str <\/span><span class=\"c4\">= Field(description=<\/span><span class=\"c5\">&#8220;Nama lengkap pengguna&#8221;<\/span><span class=\"c1\">)<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 umur: <\/span><span class=\"c21\">int <\/span><span class=\"c4\">= Field(description=<\/span><span class=\"c5\">&#8220;Umur dalam satuan tahun&#8221;<\/span><span class=\"c1\">)<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 pekerjaan: <\/span><span class=\"c13 c27\">str<\/span><\/p>\n<p class=\"c9\"><span class=\"c13\">\u00a0 \u00a0 <\/span><span class=\"c4\">skills: <\/span><span class=\"c13\">list<\/span><span class=\"c4\">[<\/span><span class=\"c13\">str<\/span><span class=\"c4\">] = Field(default_factory=<\/span><span class=\"c13\">list<\/span><span class=\"c4\">, description=<\/span><span class=\"c5\">&#8220;Daftar keahlian teknis&#8221;<\/span><span class=\"c1\">)<\/span><\/p>\n<p class=\"c8\">\n<p class=\"c9\"><span class=\"c7 c27\"># Query diproses dan di-parse langsung ke objek Pydantic<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">user: UserProfile = client.chat.completions.create(<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 model=<\/span><span class=\"c5\">&#8220;gpt-4o&#8221;<\/span><span class=\"c1\">,<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 response_model=UserProfile,<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 messages=[<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 {<\/span><span class=\"c5\">&#8220;role&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;user&#8221;<\/span><span class=\"c4\">, <\/span><span class=\"c5\">&#8220;content&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;Siti adalah DevOps Engineer usia 31 tahun yang menguasai Docker dan Kubernetes.&#8221;<\/span><span class=\"c1\">}<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 ]<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">)<\/span><\/p>\n<p class=\"c8\">\n<p class=\"c9\"><span class=\"c13\">print<\/span><span class=\"c4\">(<\/span><span class=\"c5\">f&#8221;Nama: {user.nama}, Skills: {user.skills}&#8221;<\/span><span class=\"c4\">)<\/span><\/p>\n<h3 id=\"h.thnd3pjmrgm2\" class=\"c14\"><span class=\"c23\">\uec02<\/span><\/h3>\n<h3 class=\"c14\"><span class=\"c24 c16\">3. Self-Correction Loop \/ Reflection Pattern<\/span><\/h3>\n<p class=\"c9\">Jika terjadi kesalahan parsing pada model yang tidak mendukung <span class=\"c6\">constrained decoding<\/span>\u00a0secara native, terapkan pola koreksi diri (<span class=\"c6\">Self-Correction Loop<\/span>). Ketika <span class=\"c18\">ValidationError<\/span>\u00a0tertangkap, tangkapan pesan kesalahan (<span class=\"c6\">error trace<\/span><span class=\"c3\">) dikirimkan kembali ke LLM agar model dapat memperbaiki outputnya sendiri.<\/span><\/p>\n<p class=\"c9\"><span class=\"c23\">\uec03<\/span><span class=\"c13\">from <\/span><span class=\"c4\">pydantic <\/span><span class=\"c13\">import <\/span><span class=\"c1\">BaseModel, ValidationError<\/span><\/p>\n<p class=\"c9\"><span class=\"c13\">def <\/span><span class=\"c4\">execute_agent_with_retry(prompt: <\/span><span class=\"c13\">str<\/span><span class=\"c4\">, max_retries: <\/span><span class=\"c21\">int <\/span><span class=\"c4\">= <\/span><span class=\"c20\">3<\/span><span class=\"c1\">):<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 messages = [{<\/span><span class=\"c5\">&#8220;role&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;user&#8221;<\/span><span class=\"c4\">, <\/span><span class=\"c5\">&#8220;content&#8221;<\/span><span class=\"c1\">: prompt}]<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 <\/span><span class=\"c13\">for <\/span><span class=\"c4\">attempt <\/span><span class=\"c13\">in range<\/span><span class=\"c1\">(max_retries):<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 \u00a0 \u00a0 raw_response = call_llm(messages)<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c13\">try<\/span><span class=\"c1\">:<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c7 c27\"># Uji validitas JSON dan Schema<\/span><\/p>\n<p class=\"c9\"><span class=\"c7\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c1\">validated_data = UserProfile.model_validate_json(raw_response)<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c13\">return <\/span><span class=\"c1\">validated_data<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c13\">except <\/span><span class=\"c4\">ValidationError <\/span><span class=\"c13\">as <\/span><span class=\"c1\">e:<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c7 c27\"># Masukkan kesalahan ke dalam riwayat percakapan untuk koreksi<\/span><\/p>\n<p class=\"c9\"><span class=\"c7\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c4\">messages.append({<\/span><span class=\"c5\">&#8220;role&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;assistant&#8221;<\/span><span class=\"c4\">, <\/span><span class=\"c5\">&#8220;content&#8221;<\/span><span class=\"c1\">: raw_response})<\/span><\/p>\n<p class=\"c9\"><span class=\"c1\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 messages.append({<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;role&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5\">&#8220;user&#8221;<\/span><span class=\"c1\">, <\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c5\">&#8220;content&#8221;<\/span><span class=\"c4\">: <\/span><span class=\"c5 c27\">f&#8221;JSON yang Anda berikan tidak valid. Error: {e}. Harap perbaiki output JSON Anda.&#8221;<\/span><\/p>\n<p class=\"c9\"><span class=\"c5\">\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 <\/span><span class=\"c1\">})<\/span><\/p>\n<p class=\"c9\"><span class=\"c4\">\u00a0 \u00a0 <\/span><span class=\"c13\">raise <\/span><span class=\"c4\">RuntimeError(<\/span><span class=\"c5\">&#8220;Agent gagal menghasilkan JSON valid setelah beberapa percobaan.&#8221;<\/span><span class=\"c4\">)<\/span><\/p>\n<h3 id=\"h.nbaa92lc6duw\" class=\"c14\"><span class=\"c23\">\uec02<\/span><\/h3>\n<h3 class=\"c14\"><span class=\"c24 c16\">4. Perjelas Field Description pada Tool Definition<\/span><\/h3>\n<p class=\"c9\">Kualitas prompt pada definisi alatan (<span class=\"c6\">tools<\/span><span class=\"c3\">) menentukan tingkat akurasi schema. Jangan menganggap model memahami konteks field hanya dari nama variabelnya.<\/span><\/p>\n<ul class=\"c28 lst-kix_387jvx3aio6c-0 start\">\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Hindari:<\/span>\u00a0<span class=\"c18\">&#8220;email&#8221;: {&#8220;type&#8221;: &#8220;string&#8221;}<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Gunakan:<\/span>\u00a0<span class=\"c18\">&#8220;email&#8221;: {&#8220;type&#8221;: &#8220;string&#8221;, &#8220;description&#8221;: &#8220;Alamat email resmi pengguna dengan format valid, contoh: nama@domain.com&#8221;}<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Gunakan Enum secara Eksplisit:<\/span>\u00a0Jika nilai variabel bersifat terbatas, berikan batasan enum langsung pada schema:<br \/>\n<span class=\"c18\">&#8220;status&#8221;: {&#8220;type&#8221;: &#8220;string&#8221;, &#8220;enum&#8221;: [&#8220;pending&#8221;, &#8220;approved&#8221;, &#8220;rejected&#8221;]}<\/span><\/li>\n<\/ul>\n<h2 id=\"h.vp1t1ibnb255\" class=\"c19\"><span class=\"c10\">Checklist Singkat Sebelum Deployment Tool-Calling<\/span><\/h2>\n<p class=\"c9\">Gunakan daftar verifikasi teknis berikut sebelum merilis fitur <span class=\"c6\">tool-calling<\/span>\u00a0atau <span class=\"c6\">JSON extraction<\/span><span class=\"c3\">\u00a0ke lingkungan produksi:<\/span><\/p>\n<ul class=\"c28 lst-kix_i9ioou4m02z6-0 start\">\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Aktifkan Native Strict Mode:<\/span>\u00a0Pastikan parameter <span class=\"c18\">strict: true<\/span>\u00a0aktif pada konfigurasi API provider yang mendukung <span class=\"c6\">Constrained Decoding<\/span><span class=\"c3\">.<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Validasi Runtime Pydantic\/Zod:<\/span><span class=\"c3\">\u00a0Pasang validator schema di layer aplikasi sebelum data diteruskan ke fungsi bisnis internal atau REST API.<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Isi Deskripsi Field Secara Spesifik:<\/span>\u00a0Seluruh field pada JSON schema memiliki deskripsi eksplisit, termasuk contoh format dan batasan nilai (<span class=\"c6\">enum<\/span><span class=\"c3\">).<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Terapkan Fallback &amp; Retry Handler:<\/span>\u00a0Mekanisme penanganan error (<span class=\"c6\">retry loop<\/span><span class=\"c3\">) terpasang dengan batas maksimal pengulangan 2\u20133 kali untuk mencegah boros kredit API.<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Pembersihan String Mentah (<\/span><span class=\"c16 c6\">Sanitization<\/span><span class=\"c16\">):<\/span>\u00a0Jalankan fungsi pembersih <span class=\"c6\">string<\/span>\u00a0(pembuang blok kode <span class=\"c18\">&#8220;`json<\/span>\u00a0dan <span class=\"c6\">whitespace<\/span><span class=\"c3\">\u00a0ekstra) sebagai pertahanan pertama sebelum proses parsing.<\/span><\/li>\n<li class=\"c9 c12 li-bullet-0\"><span class=\"c16\">Evaluasi Unit Test Schema:<\/span>\u00a0Buat <span class=\"c6\">test suite<\/span>\u00a0yang menguji ketahanan agent terhadap <span class=\"c6\">edge cases<\/span><span class=\"c3\">\u00a0seperti input pengguna yang ambigu atau kosong.<\/span><\/li>\n<\/ul>\n<p>Error parsing JSON biasanya pertama kali muncul saat agent mulai memanggil tool. Alur dasarnya kami bahas di <a href=\"https:\/\/komunitech.com\/blog\/ai-agent\/cara-menghubungkan-ai-agent-ke-tools-api-eksternal-tutorial-kursus-gratis-2026\/\">tutorial menghubungkan AI agent ke tools dan API eksternal<\/a>, sedangkan pesan error di luar urusan JSON, seperti 401 atau 429, dijelaskan di <a href=\"https:\/\/komunitech.com\/blog\/tutorial\/arti-kode-error-pada-ai-agent-dan-solusinya\/\">panduan arti kode error pada AI agent<\/a>.<\/p>\n<h2>FAQ Seputar Parsing Error &amp; Schema JSON<\/h2>\n<h3>Mengapa LLM masih sering menyisipkan backtick markdown (&#8220;`json) padahal sudah dilarang di prompt?<\/h3>\n<p>Karena data pre-training LLM didominasi oleh teks percakapan dan dokumentasi teknis yang terbiasa memformat kode dengan markdown. Jika hanya mengandalkan prompt, probabilitas token markdown tetap ada. Solusi paling pasti adalah menggunakan API Structured Outputs native (strict mode) atau regex\/stripper di layer kode backend sebelum parsing.<\/p>\n<h3>Apakah Pydantic otomatis membetulkan tipe data yang salah dari LLM?<\/h3>\n<p>Ya, Pydantic memiliki mekanisme type coercion. Misalnya, jika LLM mengembalikan string &#8217;42&#8217; untuk field berjenis integer, Pydantic akan otomatis mengonversinya menjadi angka 42. Namun, jika LLM mengembalikan nilai yang tidak bisa dikonversi (misal: &#8216;empat puluh dua&#8217;), ValidationError akan dilempar.<\/p>\n<h3>Kapan perlu memakai Self-Correction Loop vs Native Structured Outputs?<\/h3>\n<p>Jika provider LLM yang kamu pakai mendukung native structured outputs (seperti OpenAI Strict Mode atau Gemini JSON Mode), prioritaskan fitur native tersebut karena tingkat keberhasilannya 100% tanpa boros token. Gunakan Self-Correction Loop sebagai fallback saat memakai open-source LLM lokal (Ollama\/vLLM) yang belum mendukung constrained decoding tingkat server.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Membangun AI Agent yang andal untuk lingkungan produksi sering kali berbenturan dengan kenyataan pahit: Large Language Models (LLM) pada dasarnya adalah mesin pemprediksi token berturut-turut, bukan parser sintaksis yang deterministik. Saat kita meminta AI Agent memanggil alat (tool-calling) atau menghasilkan payload JSON untuk integrasi sistem, ketidakstabilan format sering menjadi batu sandungan utama. Apa itu Parsing [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":2943,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[74],"tags":[],"class_list":["post-2942","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-karir-skill"],"acf":[],"_links":{"self":[{"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/posts\/2942","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/comments?post=2942"}],"version-history":[{"count":3,"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/posts\/2942\/revisions"}],"predecessor-version":[{"id":2992,"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/posts\/2942\/revisions\/2992"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/media\/2943"}],"wp:attachment":[{"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/media?parent=2942"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/categories?post=2942"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/komunitech.com\/blog\/wp-json\/wp\/v2\/tags?post=2942"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}