Domain-Constrained Retrieval-Augmented Generation System (RAG) for Website Chatbots: The Tech Thinker Case Study

Small and medium websites increasingly require intelligent chat interfaces that can answer user questions accurately without relying on generic large language models (LLMs). However, when LLMs operate without domain boundaries, they often produce hallucinated responses that do not reflect the website's actual content. This work presents a complete, domain-constrained Retrieval-Augmented Generation (RAG) system designed specifically for The Tech Thinker, a niche engineering and AI knowledge hub with more than 100 live pages. The system combines structured web extraction, canonical URL filtering, token-based chunking, OpenAI embeddings, and a Pinecone-based vector index to build a clean and verifiable knowledge base. A lightweight routing layer and a persona-aware chat engine ensure grounded responses with a single, reliable source citation. The proposed pipeline was implemented end-to-end in seven steps and evaluated using a 15-query benchmark covering identity, contact, topic-coverage, technical explanations, and unsupported requests. The chatbot achieved zero hallucinations, consistent grounding, and correct routing in all cases. This study demonstrates that a small website can achieve enterprise-level answer accuracy using a reproducible, domain-restricted RAG framework with minimal infrastructure and fully transparent behavior.