4 Talking with your documents
This chapter covers
- Retrieval Augmented Generation (RAG)
- Enabling a vector store
- Creating a document loading pipeline
Thinking back on your school days, do you recall ever being told that the exam you’d be taking would be an open-book exam? No matter how much you learned or how hard you could cram for a test, knowing that you’d have the source material at your fingertips to help you answer the questions on the exam gave you greater confidence that you’d be able to answer the questions correctly.
Now imagine that instead of just being told you could use your book to look up the answers, you were told specifically which pages the answer could be found on. Being equipped with the knowledge of exactly where to look in the book would all but guarantee success.
LLMs are trained with an immense amount of information, but there are often questions that their training will not have prepared them for. Being able to pair a question with a document that includes the answer—or better yet, a small chunk of a document with the required information—can not only help the LLM answer questions more accurately, but also virtually eliminate hallucination that come about when it tries to answer questions that exceed the limits of its training.
In this chapter, you’ll look at Retrieval Augmented Generation (RAG), a way to provide relevant information to the LLM on-the-fly as you are asking questions. Let’s start by getting to know how RAG works.