RAG sends the retrieved information to a generative model to construct a specific response. A typical example is an SEO consultant summarizing documents relating to a common topic.
But let’s say you’re planning a wedding for 100 people with constraints like a vegan/vegetarian diet or allergies.
Once on your favourite e-commerce website, you want to retrieve products from these high-level specifications, not the other way around like a RAG would.
This is because you don’t want to search for dozens of products and add them to your cart.
Instead, you want the website to find relevant products according to your specifications and add them to your cart.
This would be “Fetching Instructions”.
To be honest, this is similar to OpenAI functions, or more generally AI agents: calling subsystems like search to execute a set of instructions.