RAG Research Assistant

Built during my internship on retrieval-augmented generation (RAG).

Problem. Large language models answer fluently but can make things up, and they don't know what is inside your own documents.

Approach.

• Extract text from PDFs and split it into overlapping chunks
• Embed the chunks and store them in a vector index
• Retrieve the most relevant chunks for each question
• Generate an answer that only uses the retrieved text and cites it

What I learned. Chunk size and overlap matter as much as the model; evaluating retrieval separately from generation makes problems much easier to find.