Retrieval-Augmented Generation
An approach that retrieves external documents at generation time and grounds the language model's response in them. Avoids retraining for freshness or proprietary knowledge — and surfaces citation authenticity as a new problem.
Definition
A standard RAG pipeline has four stages: (1) embed the query, (2) retrieve relevant documents via vector search, (3) concatenate the retrieved documents into the prompt, (4) the model responds grounded in those documents. Meta crystallized the pattern in 2020; it has been the industrial baseline since.
The upside is being able to address fresh information and organization-specific knowledge without retraining, and to surface source citations alongside answers. The downside is undetectable failure when retrieved documents are tampered with or citations are fabricated.
Running RAG in a regulated context requires proving both that the retrieved corpus has a trustworthy lineage and that the citations in the response actually come from those documents. This is the boundary at which verifiable AI becomes necessary.
Lemma Oracle implementation
Lemma layers three guarantees onto a RAG pipeline: (1) the retrievable corpus is lineage-pinned by CID and docHash; (2) retrieval results carry provenance metadata; (3) citation proofs bind each cited string to its source document.
The recipient can verify, without re-receiving the content, that the answer's text really came from the cited document, that the document came from a trusted issuer, and that the document has not been tampered with.
Financial research assistants, medical decision support, legal AI case-law citation — any domain where citation authenticity bears on professional responsibility — gets a concrete solution out of this stack.