{"id":7789,"date":"2026-05-26T09:00:00","date_gmt":"2026-05-26T01:00:00","guid":{"rendered":"https:\/\/drhariz.com\/blog\/?p=7789"},"modified":"2026-05-21T18:01:47","modified_gmt":"2026-05-21T10:01:47","slug":"rag-bahasa-melayu","status":"publish","type":"post","link":"https:\/\/drhariz.com\/blog\/rag-bahasa-melayu\/","title":{"rendered":"RAG for Bahasa Melayu: Building Retrieval-Augmented Generation in Local Language"},"content":{"rendered":"

What is RAG for Bahasa Melayu and why does Malaysia need it?<\/h2>\n

RAG Bahasa Melayu is retrieval-augmented generation that combines a large language model with a Malay-language knowledge base to produce accurate, locally relevant answers. Malaysia needs it because off-the-shelf models often hallucinate on local laws, policies, and cultural context, while RAG grounds outputs in trusted Malay sources.<\/p>\n

Dr. Muhamad Hariz Muhamad Adnan, an HRD Corp Certified AI<\/a> Trainer at Universiti Pendidikan Sultan Idris (UPSI<\/a>), works with Malaysian agencies and SMEs to deploy RAG systems that answer customer and citizen queries fluently in Bahasa Melayu without expensive model fine-tuning.<\/p>\n

How does a RAG Bahasa Melayu system work end-to-end?<\/h2>\n

A RAG Bahasa Melayu system works by chunking Malay documents, converting them to vector embeddings, retrieving the most relevant chunks for a user query, and passing them as context to an LLM. The LLM then generates a grounded Malay-language answer, citing the retrieved sources for transparency and trust.<\/p>\n

The five-step RAG pipeline<\/h3>\n
    \n
  1. Ingest:<\/strong> Collect Malay PDFs, policies, FAQs, and websites.<\/li>\n
  2. Chunk:<\/strong> Split documents into 300\u2013800 token segments.<\/li>\n
  3. Embed:<\/strong> Use a multilingual embedding model that handles Bahasa Melayu well.<\/li>\n
  4. Retrieve:<\/strong> Query the vector database for top-K relevant chunks.<\/li>\n
  5. Generate:<\/strong> Pass chunks plus user question to GPT-4o, Claude, or Gemini.<\/li>\n<\/ol>\n

    Which embedding models handle Bahasa Melayu best?<\/h2>\n

    The strongest embedding models for Bahasa Melayu in 2026 are OpenAI text-embedding-3-large, Cohere embed-multilingual-v3, BGE-M3, and Google Vertex AI multilingual embeddings. These models capture Malay morphology, code-switching with English, and Malaysian named entities better than English-only models like Ada-002 or older sentence-transformers.<\/p>\n\n\n\n\n\n\n\n
    Embedding Model<\/th>\nBahasa Melayu Quality<\/th>\nCost per 1M Tokens (USD)<\/th>\nOpen Source<\/th>\n<\/tr>\n
    OpenAI text-embedding-3-large<\/td>\nExcellent<\/td>\n0.13<\/td>\nNo<\/td>\n<\/tr>\n
    Cohere embed-multilingual-v3<\/td>\nExcellent<\/td>\n0.10<\/td>\nNo<\/td>\n<\/tr>\n
    BGE-M3<\/td>\nVery good<\/td>\nSelf-hosted<\/td>\nYes<\/td>\n<\/tr>\n
    Google Vertex multilingual<\/td>\nVery good<\/td>\n0.025<\/td>\nNo<\/td>\n<\/tr>\n
    Ada-002 (legacy)<\/td>\nWeak<\/td>\n0.10<\/td>\nNo<\/td>\n<\/tr>\n<\/table>\n

    What are the best use cases for RAG Bahasa Melayu in Malaysia?<\/h2>\n

    The top RAG Bahasa Melayu use cases are government citizen services, university student support, Islamic finance compliance, healthcare patient triage, and SME customer service. These domains demand accurate, source-cited answers in Malay, which RAG provides without the cost or risk of fine-tuning a Malay-specific model from scratch.<\/p>\n