# Building a 100% Free RAG Digital Twin with Gemini 2.0 & MySQL Have you ever wanted to talk to a PDF? Or better yet, have an AI version of yourself talk to your recruiters based on your actual career documentation? In this guide, I'll walk you through how I built my "Digital Twin" chatbot using a high-performance **RAG (Retrieval-Augmented Generation)** architecture—all while keeping it 100% free using Google's Gemini 2.0 Flash and a standard MySQL database. --- ## The Architecture: How it Works Most AI chatbots "guess" based on general training. A Digital Twin must be grounded in facts. Here is the 4-step pipeline I implemented: ### 1. The Ingestion Engine (PDF to Knowledge) We don't just "upload" a PDF. We use a custom Node.js script and `pdfjs-dist` to extract raw text. But raw text is messy. To make it AI-ready, I implemented **Semantic Chunking**: * **Sentence-Aware Splitting**: We break the text into 1,000-character "nuggets." * **Overlapping Context**: Each nugget starts with the previous sentence so the AI never loses the thread of a thought. * **1st-Person Transformation**: During ingestion, we use Regex to convert 3rd-person mentions ("Ashish did X") into 1st-person facts ("I did X"). ### 2. The Storage (MySQL Full-Text Search) Instead of expensive vector databases like Pinecone, I utilized **MySQL Full-Text Indexing**. By adding a `FULLTEXT` index to the `content` column, we can perform "Natural Language" queries at lightning speed. ```sql SELECT content FROM ai_knowledge WHERE MATCH(content) AGAINST ('How much experience does Ashish have?' IN NATURAL LANGUAGE MODE); ``` ### 3. The Brain (Gemini 2.0 Flash) When a user asks a question, we don't send the whole PDF to the AI (that would be slow and expensive). Instead: 1. We find the top 3 most relevant "nuggets" in the DB. 2. We send **only those nuggets** + the user question to **Gemini 2.0 Flash**. 3. We use a "Persona-Locked" prompt to ensure the AI speaks only as Me. ### 4. The Voice (Web Speech API) To make the twin truly "alive," I integrated the browser's `SpeechSynthesisUtterance` API. The moment the AI synthesizes an answer, the browser speaks it back in a natural human voice. --- ## Why this is a Game Changer for Portfolios * **Zero Cost**: Uses Gemini's generous free tier and your existing hosting DB. * **Highly Scalable**: You can add 100 PDFs or 100,000 nuggets, and MySQL will still find the answer in milliseconds. * **Persona Integrity**: By transforming the data during ingestion, the AI never breaks character. This project proves that you don't need a massive budget to build state-of-the-art AI features. You just need the right architecture.