RAG gets described like a single feature: 'give the model documents'. In practice, it is a pipeline with separate failure modes.
You have a knowledge source, a way to chunk it, an embedding or retrieval strategy, a context selection step and finally a generation step. Each layer changes the quality of the answer.
Retrieval is the first quality gate
If the right information never reaches the model, a better prompt cannot repair the missing context. That is why retrieval quality deserves its own tests rather than being treated as plumbing.
The model should not be the database
A model is good at language generation. It is not your source of truth for institution-specific or frequently changing data. RAG lets an application bring external evidence into the generation step.
Context is a budget
More retrieved text is not automatically better. Context has cost, latency and attention limits. A good system retrieves enough evidence to answer the question without flooding the prompt with irrelevant material.
Then comes the agent
An agent adds another decision: which tool to call. In the admission-assistant experiment in this portfolio, the agent can use a dedicated RAG tool and a search tool, then answer under a tightly constrained response policy.
Related case studies