Smart RAG
June 2026
A full-stack Retrieval-Augmented Generation app that lets users chat with their own PDF documents. Uploaded files are parsed, chunked and embedded, and every question is answered strictly from the most relevant passages. A dedicated LLM security judge screens each message first, blocking prompt-injection, jailbreak and off-topic requests before they reach the answering model.
Key Features
- PDF upload with automatic text extraction and overlapping chunking (1000 chars, 200 overlap)
- Semantic search over chunk embeddings (Google text-embedding-004) ranked by cosine similarity
- Two-stage LLM pipeline on Groq: a security judge model gates requests before the answer model runs
- Grounded answers that cite the source passages and refuse to answer outside the document context
- Conversation memory for follow-up questions, plus an in-app PDF viewer and document library
- JWT authentication with per-user document isolation in MongoDB
Tech Stack
ReactTailwindFramer MotionNode.jsExpress.jsMongoDBGroqJWT