Understanding RAG applications – closed-book versus open-book question-answering

We are all used to modern search engines. They solve our information retrieval needs by making us follow a common pattern [3]:

Figure 3.1: A typical flow that fulfills information retrieval needs

Without going into too much detail, on the architectural level, model information retrieval systems use a corpus (that is typically some kind of inverted index). They then use ranking and retrieval to find the information to answer the query:

Figure 3.2: A typical flow that fulfills information retrieval needs

Already in 2020, it was demonstrated that LLMs do memorize facts they have seen during training, and they can be used as knowledge repositories [16]. In 2021, a model-based approach was suggested for search and question-answering tasks [2]. There are two options for how we can use LLMs. First, we can fine-tune an LLM and use it to complete...