📚 Smart PDF RAG System

An intelligent, stateful web application that allows users to upload any PDF document and instantly interact with its contents. This system leverages a Retrieval-Augmented Generation (RAG) architecture to completely eliminate LLM hallucination by grounding the AI's responses strictly in the provided document context.

🛠️ Technical Architecture

• Frontend UI: Streamlit (with Stateful Chat Memory)
• LLM Orchestration: LangChain
• Inference Engine: Groq API (Ultra-low latency)
• Generative Model: LLaMA 3.3 70B Versatile
• Embedding Model: Hugging Face (sentence-transformers/all-mpnet-base-v2)
• Vector Database: ChromaDB (Ephemeral, in-memory processing)
• Document Parsing: PyPDFLoader

⚙️ Core System Features

1. Dynamic Document Ingestion & Processing

• Accepts user-uploaded PDFs via Streamlit's in-memory file uploader.
• Utilizes secure OS-level temporary file generation (tempfile) to bridge the gap between Streamlit's byte-stream and LangChain's file-path requirements.
• Automatically cleans up temporary storage immediately after processing to prevent memory leaks.

2. Optimized Text Chunking

• Implements RecursiveCharacterTextSplitter with a strict chunk_size=1000 and chunk_overlap=100.
• This specific overlap ensures that contextual meaning is not lost across chunk boundaries, resulting in highly accurate embeddings.

3. Localized Vector Search (MMR)

• Uses a localized Hugging Face sentence-transformer model to generate dense vector embeddings, reducing API costs and latency.
• Stores embeddings in a ChromaDB vector store.
• Utilizes Maximum Marginal Relevance (MMR) search (lambda_mult=0.5) to retrieve the top 4 chunks that are not only highly relevant to the user query but also diverse from each other.

4. Stateful UI & Caching

• Leverages st.session_state to cache the instantiated retriever engine. This ensures the computationally expensive embedding process only occurs once per document upload, rather than refreshing on every user interaction.