Evaluating embedding tasks

- Back to our tree of LLM evaluations. We are moving on beyond generative tasks to understanding tasks. Arguably, the harder concept to really wrap your head around, because at least with generative tasks, we are able to look at what our LLMs are saying and ideally check if they are correct or not. But we have some cohesive, coherent way of, as a human, gut checking the response. We can read it. With understanding tasks, this becomes more complicated and we have to rely a little bit more on mathematical evaluations than human-based evaluations. Let's start with embeddings. Well, what are embeddings? Embeddings are some way of representing data. Now, in our case, we'll be mostly talking about text embeddings, but really, any kind of embedding will take data and convert them to numerical vectors, which are one-dimensional (indistinct) series of numbers, such that data that are related in some way result in vectors that are near each other. Retrieval is one of the most common ways we use…