Glossary entry · ai term
RAG
Retrieval-augmented generation — pattern that grounds an LLM's output in retrieved source documents to reduce hallucination.
Why it matters in tax practice
Pure LLMs hallucinate citations — a known and documented failure mode in tax research. RAG architectures retrieve from primary sources (statute, regs, IRS guidance) and pass the retrieved text to the model, which reduces (does not eliminate) hallucination on citation-heavy work. Blue J and CoCounsel are the named statute-anchored tools; consumer LLMs are not RAG-grounded by default. Practitioner due-diligence under Circular 230 §10.22 still requires verifying every citation against the canonical source.