Retrieval and RAG: useful, leaky if unbounded

Retrieval makes answers better by fetching context. Unbounded retrieval makes the model a search engine over everything it should never have seen.

RAG is not magic memory. It is fetch-then-generate: index, query, top-k chunks, then a prompt. Usefulness and leakiness share the same pipe.

Each stage can over-share. An over-broad corpus puts neighboring secrets next to the answer. A loose query pulls more than the task needs. Chunks can include paths, headers, or ACL-hidden neighbors the UI never showed. Citations can reveal where sensitive files live even when the prose looks harmless.

Bound three things or you rebuilt shadow search: the corpus (what is eligible), the query (what may be asked), and the returned chunks (what may enter the prompt). Identity-aware retrieval, redaction before embed, and a kill-switch on indexes are controls — not a product pitch.

Retrieved context is still a data path. Sometimes it is safer than pasting a whole file into the box. It is never free. The same questions from the data-path field note apply: what left, what persisted, who can trigger the next hop, can you revoke mid-path.

This piece explains retrieval as AI × Data Security. It does not sell a suite. Named platforms appear later only with Learn-backed facts.