LLM Fundamentals

An LLM fundamental, in this path, is a piece of the application you can draw on one page and still operate next month. This sub-branch starts with retrieval-augmented generation because that is the first design teams ship, and it continues with evaluation because a pipeline without a check becomes a demo that cannot be changed safely.

You should know what a request is and what a database query returns. You do not need a research background. The posts avoid leaderboard talk except where a benchmark is being misused as a product metric. The useful split is simple: a benchmark tells you how a model did on a task someone else defined, and an eval tells you whether your application did the job your user asked for.

Read the retrieval post before the evaluation post. The second one keeps referring to stages of a pipeline, and those stages are named in the first. If you already run a retrieval system, the evaluation note is the one that changes how you review a prompt diff.

Both pages are budding. I use these designs, and I still expect to revise the failure cases as I see new ones in production. The tended date is the honest signal of how recently that happened. There is no exercise platform here. The exercise is to point the pipeline at a corpus you actually have and to write down what a wrong answer costs.