Here's how it works: The PDF document is divided into smaller sections, and each section is stored in a database with its corresponding embedding vector. When users ask questions, a similarity search is performed against the embedding vectors to retrieve relevant sections.