Words
Retrieval finds text. It does not make that text safe.
Retrieval selects information from stored sources for a task. A query is the request used to find it. A search index organizes information so it can be found. A cache stores a copy for speed.
A chunk is a piece of a document handled on its own. An embedding is a numeric stand-in used to compare content. A vector store keeps those numbers, and a similarity search finds nearby ones. A result can be relevant and still be false, unsafe, or forbidden. Search quality and permission are different checks.
Retrieval-augmented generation means the system fetches text and puts it into the model’s input. It does not usually mean the model was retrained on that text. Derived data is made from other data: summaries, embeddings, extracted fields, cached snippets. A derived copy can still be sensitive.
A log records events. A trace connects the steps of one request. A backup is a copy kept for recovery. A secret, such as an API key, does not belong in a prompt, a summary, an error, or a general file the agent can read. Data minimization means using only the fields the task needs.