RAG PDF Chatbot with Phi-2/TinyLlama

A full-stack RAG (Retrieval-Augmented Generation) application built to process and query unstructured PDF data. The system implements a sophisticated pipeline: documents are partitioned into semantic chunks, transformed into high-dimensional embeddings, and indexed in a FAISS vector store. At query time, the system performs semantic retrieval to provide grounded context to local Small Language Models (TinyLlama or Phi-2), ensuring accurate answers while maintaining 100% data privacy.

Key Features

• End-to-End RAG Pipeline: Seamless integration of document loading, semantic chunking, and vector indexing.
• Semantic Vector Search: Utilizes FAISS and Sentence-Transformers for high-speed, context-aware information retrieval.
• Local SLM Integration: Optimized for privacy and performance using TinyLlama and Microsoft's Phi-2 models via HuggingFace.
• Explainable AI (XAI): Features a source-tracking mechanism that highlights exactly which document segments were used to generate each answer.
• Interactive UI: A polished Streamlit-based chat interface with session persistence and real-time latency monitoring.
• Optimized Text Processing: Implements Recursive Character Splitting with overlap to preserve semantic continuity across chunks.

Tech Stack

Screenshots