Enterprise Document Intelligence builds a RAG system from four bricks: parsing, question analysis, retrieval, and generation. This article turns attention to the fourth brick’s after‑generation phase, where a validator checks the model’s answer for shape, evidence, and format before it reaches the user.\n\nThe validator first confirms that the answer matches the expected schema chosen for the brief. It then verifies that every span references a real line in the source PDF, that each quoted phrase is a substring of the cited lines after whitespace and reference normalisation, and that dates and currencies follow ISO 8601 and ISO 4217 standards.

Cross‑field checks enforce logical consistency, such as ensuring a contract’s start_date precedes its end_date and that summed line items match a declared total.\n\nVerbatim extraction proves fragile: models may output quotes with altered spacing, strip bibliographic references like [9], or misalign line numbers by one or two. The validator’s substring plus normalisation logic catches these subtle mismatches that would otherwise slip through, preserving semantic fidelity for legal or compliance contexts.\n\nWhen validation fails, the pipeline can retry with a stricter prompt, flag the result for human review, or reject it outright. Choosing the correct response depends on the risk level; audit‑critical paths often reject any non‑conforming answer to maintain integrity.