Engineered RAG-Based Conversational PDF System with LangChain

Developed a retrieval-augmented generation (RAG) conversational pipeline in Python using LangChain, text splitters, vector embeddings, and large language models to enable semantic question-answering over uploaded PDF documents.

Implemented document ingestion, semantic chunking, vector store indexing, and contextual query prompting.

Chunking strategy and overlap size are the single most decisive factor in RAG accuracy. Splitting text too finely loses contextual meaning, while chunks that are too large dilute semantic embedding similarity.

Add re-ranking (Cross-Encoder) on retrieved chunks before passing context to the LLM to filter out borderline embeddings and improve answer precision.