Lingua RakyatLingua Rakyat

Guidance

Read official documents in language that feels clearer.

Malaysian MyKad identity card
Malaysian passport
Malaysian civic guide helping users understand official documents
Multilingual AI for Government Documents

Government documents, explained inyour language.

An AI-powered government document Q&A system using Cohere multilingual embeddings, Cohere neural reranking, and LLM. Source-cited answers with faithfulness scoring in Malay, English, and Chinese.

Semantic Search
ROUGE-1/2/L
Fast Inference
Confidence Scoring
Bahasa Melayu • English • 中文 •
Multilingual RAG • Civic AI •
BANTUAN_RAKYAT_1MALAYSIA_2024.PDF
> 
According to Section 2.1, applicants must be Malaysian citizens aged 18 and above with a household income ≤ RM4,000.
— Chunk ID: doc_001_chunk_42 | Similarity: 0.89
>

01Differentiation vs Generic RAG

Built for Malaysian civic AI, not generic chatbots.

Not "GPT + PDF upload". Positioned as "Malaysian-first civic AI": speaks Malay naturally, knows government context, and stays transparent with sources + pages.

FeatureLingua RakyatChatGPT + PDF upload
Natural Malay supportyesyes
Malaysia government-firstyes—
PDF citations + pagesyesyes
Confidence badgeyes—
Offline mode (cache)yes—
Multilingual semantic retrievalyesyes
Document managementyes—
Built-in eval metrics (ROUGE/BLEU)yes—
Faithfulness scoring (answer vs sources)yes—
Voice I/O (STT + TTS)yesyes

Note: Offline mode depends on documents and excerpts cached earlier while online.

02System Flow

01

Ingestion

PDF → text extraction → section-aware chunking (360-word window, 45-word overlap) → Cohere embed-multilingual-v3 → Pinecone.

02

Retrieval

Language detection → multi-query augmentation (×4 variants) → Cohere embedding → Pinecone vector search → Cohere neural reranking → top-k chunks.

03

Generation

Evidence guard → context-aware prompt → LLM → streamed answer + source citations + confidence + faithfulness score.

03Key Features

A complete RAG system with an evaluation dashboard for transparency and reliability.

Multilingual RAG

Cohere embed-v3 with support for 100+ languages, generating outputs in BM, EN, or zh.

Evaluation Suite

ROUGE, BLEU, Flesch-Kincaid, exact match, and real-time metrics.

Voice I/O

Ask by voice (Whisper STT) and hear answers read back (ElevenLabs TTS). Browser speechSynthesis fallback when offline.

Evidence Guard

Refuses to hallucinate — only answers when retrieved evidence clears the confidence threshold. Every answer carries a faithfulness score.

04Technical Demonstration

System processing real government documents

05Technology Stack