ExploreSingapore

ven't build one yet or building one and are stuck somewhere also Singaporean people(obviously!) Comparison:- RAW LLM vs RAG based LLM to test the rag implementation i compared output of my logic code against the standard(gemini/Arcee AI/groq) and custom system instructions with rag(gemini/Arcee AI/groq) results were shocking query:- "can I fly in a drone in public park" standard llm response :- ""gave generic advice about "checking local laws" and safety guidelines"" Customized llm with RAG :- ""cited the air navigation act,specified the 5km no fly zones,and linked to the CAAS permit page"" the difference was clear and it was sure that the ai was not hallucinating. Ingestion:- I have the RAG Architecture about 594 PDFs about Singaporian laws and acts which rougly contains 33000 pages. How did I do it :- I used google Collab to build vector database and metadata which nearly took me 1 hour to do so ie convert PDFs to vectors. How accurate is it:- It's still in development phase but still it provides near accurate information as it contains multi query retrieval ie if a user asks ("ease of doing business in Singapore") the logic would break the keywords "ease", "business", "Singapore" and provide the required documents from the PDFs with the page number also it's a little hard to explain but you can check it on my webpage.Its not perfect but hey i am still learning. The Tech Stack: Ingestion: Python scripts using PyPDF2 to parse various PDF formats. Embeddings: Hugging Face BGE-M3(1024 dimensions) Vector Database: FAISS for similarity search. Orchestration: LangChain. Backend: Flask Frontend: React and Framer. The RAG Pipeline operates through the following process: Chunking: The source text is divided into chunks of 150 with an overlap of 50 tokens to maintain context across boundaries. Retrieval: When a user asks a question (e.g., "What is the policy on HDB grants?"), the system queries the vector database for the top k chunks (k=1). Synthesis: The system adds these chunks to the prompt of LLMs which produces the final response that includes citation information. Why did I say llms :- because I wanted the system to be as non crashable as possible so I am using gemini as my primary llm to provide responses but if it fails to do so due to api requests or any other reasons the backup model(Arcee AI trinity large) can handle the requests. Don't worry :- I have implemented different system instructions for different models so that result is a good quality product. Current Challenges: I am working on optimizing the the ranking strategy of the RAG architecture. I would value insights from anyone who has encountered RAG returning unrelevant documents. Feedbacks are the backbone of improving a platform so they are most Repository:- https://github.com/adityaprasad-sudo/Explore-Singapore

I use a "Triple Failover" system with three models as backup. When the main model failed, the backups sounded entirely different. The Solution: I added Dynamic System Instructions. Now, if the backend switches to Model B, it uses a specific prompt designed for Model B’s features, making it mimic the structure and tone of the primary model. The user never notices the change. 2. The "Deep Search" Fix Issue: A simple semantic search for "Starting a business" misses related laws like "Tax" or "Labor" acts. The Solution: I implemented Multi-Query Retrieval (MQR). An LLM now intercepts your query. It breaks it down into sub-intents (e.g., “Business Registration,” “Corporate Tax,” “Employment Rules”). It searches for all of them at the same time and combines the results. Result: Much richer, context-aware answers. 3. The "Hallucination" Fix Issue: Garbage In, Garbage Out. If FAISS retrieves a bad document, the LLM produces inaccurate information. The Solution: I added a Cross-Encoder Re-Ranking layer. Step 1: FAISS grabs the top 10 results. Step 2: A specialized Cross-Encoder model evaluates them for relevance. Step 3: Irrelevant parts are removed before they reach the Chat LLM. * The Tech Stack * Embeddings: BGE-M3 (Running locally) Vector DB: FAISS Backend: Python + Custom Triple-Model Failover Logic: Multi-Query + Re-Ranking (New in V2) Try it out I am still learning. I’d love to hear your thoughts on the new logic. Live Demo: https://adityaprasad-sudo.github.io/Explore-Singapore/ GitHub Repo: https://github.com/adityaprasad-sudo/Explore-Singapore Feedback, especially on the failover speed, is welcome!