Language and retrieval systems

Retrieval and grounded generation

Retrieval-augmented generation fetches relevant material from a controlled source before a language model answers, so responses are grounded in your documents rather than in the model's general training.

Operational use
Answering operational questions from policy documents, contracts, ledgers, or procedure libraries, with the source cited beside the answer.
When not to use it
When the source corpus is thin, stale, or wrong; retrieval faithfully grounds answers in bad material. Also not a fit when a deterministic lookup would do.
How it is validated
Evaluate on a held-out set of real questions with known answers; measure retrieval hit rate separately from answer quality; require citations that a reviewer can check.
Limitations
Retrieval quality, source quality, permissions, and answer evaluation remain separate problems. Transformers made these systems possible; they did not make outputs automatically factual.

The transformer architecture is the foundation of modern language and multimodal models. It made large-scale language systems practical, but nothing about the architecture guarantees truthfulness: a model generates what is likely, not what is verified.

Retrieval-augmented generation narrows that gap by fetching relevant passages from a controlled corpus before generation, and asking the model to answer from them. In operational use, this is the difference between “what does the model believe” and “what do our documents say.”

Eigenvalue treats four things as separate engineering problems: what gets into the corpus, how well retrieval finds the right passages, who is permitted to see what, and whether the final answer is actually supported by the retrieved text. A system is only as trustworthy as the weakest of the four.

Where it appears in our work

Primary sources

Last reviewed 2026-08-02. Sources link to primary or official pages; verify current versions before relying on them.