GitHub - adityaprasad-sudo/ExploreSingapore: A sophisticated RAG intelligence engine for Singaporean laws, policies, and history. Comes with a triple-AI failover backend (Gemini/Llama/Groq), semantic embeddings using FAISS, and an Apple-inspired interactive UI. Designed with precision and high availability in mind.

2 min read Original article ↗

Explore Singapore

Legal, Historical, and Infrastructural Knowledge Engine

Python Flask LangChain FAISS HuggingFace webpage screenshot

📌 Project Overview

The Singapore Intelligence RAG System is an intelligent platform that utilizes AI technology to deliver accurate and relevant information about the legal system, policies, and historical events of Singapore, as well as its critical infrastructure.

Unlike other LLMs, which have the tendency to "hallucinate" facts, the Singapore Intelligence RAG System employs Retrieval-Augmented Generation (RAG). It relies on a carefully curated set of Singaporean data (more than 33,000 pages of PDFs) to ensure that all answers are based on factual reality.

🏗 System Architecture

The system follows a high-performance RAG pipeline optimized for low-resource environments:

  1. Ingestion: Processed 33,000+ pages of Singaporean legal and historical documents.
  2. Vectorization: Used BGE-M3 to create 1024-dimensional semantic embeddings.
  3. Retrieval: Implemented FAISS (Facebook AI Similarity Search) for millisecond-latency vector lookups.
  4. Generation: A "Triple-Failover" logic ensures 99.9% uptime.

🚀 Key Features

1. Triple-AI Failover Backend

For reliability in demos and heavy traffic, the system establishes a robust chain of command for LLM inference as follows:

Primary: Google Gemini 2.0 Flash (Fastest, High Context)

Secondary: Llama 3.3 70B via OpenRouter (Robust fallback)

Tertiary: Llama 3.3 70B via Groq (Emergency fallback)

2. "Lquid-Glass" Interactive UI

The frontend interface is a custom-built Framer Code Component (React + Framer Motion).

  • Glassmorphism: Real-time backdrop blur (backdrop-filter: blur(25px)).
  • Spring Physics: Smooth sideways expansion on hover.
  • Minimalist Design: SVG iconography and San Francisco typography.

3. Local Embedding Inference

Rather than using API calls for vectorization (which incurs latency and expense), the embedding model is executed locally within the application container for privacy and performance.

🛠 Tech Stack

Component Technology Description
Frontend React, Framer Motion Interactive "Ask AI" widget.
Backend Flask, Gunicorn REST API handling RAG logic.
Vector DB FAISS (CPU) Local, high-speed similarity search.
Embeddings Sentence-Transformers BGE-M3 (Local(server based)).
LLMs Gemini 2.5 flash, Llama 3.3 Text generation and synthesis.
Deployment Hugging Face Spaces Docker-based cloud hosting.

⚙️ Installation & Local Setup

Prerequisites

Very Important Install this In the backend Server before executing any Python files in the server.

  • flask
  • flask-cors
  • python-dotenv
  • google-generativeai
  • google-genai
  • langchain
  • langchain-google-genai
  • langchain-community
  • langchain-huggingface
  • faiss-cpu
  • sentence-transformers
  • pypdf
  • tiktoken
  • numpy
  • gunicorn
  • setuptools
  • wheel
  • scikit-learn
  • openai

Singapore Intelligence System Architecture

1. Clone the Repository

git clone [git clone https://github.com/adityaprasad-sudo/Explore-Singapore.git)