Large language models deployed in document processing and relational query systems exhibit two structurally distinct hallucination failure modes that existing mitigation approaches cannot resolve. In document extraction, the LLM generates plausible entities with no textual grounding in the source document. In natural language to SQL translation, the model generates queries from schema alone, without knowledge of real data distributions, producing syntactically correct SQL that returns empty or incorrect results. Both failure modes share a common property: they are not correctable with better models or better prompts because they are structural properties of the generation process itself. This paper presents two independent AI methods that address each failure mode at its root. VERA (Verified Extraction with Region Anchoring) installs a deterministic post-generation verification gate: for each LLM-generated claim, VERA identifies its source region in the document and applies a multi-technique verification algorithm — semantic similarity, edit distance, numeric normalization, and substring anchoring — against a ground-truth channel structurally independent of the generative process. VERA achieves over 99% accuracy on digital documents, 92% on cursive manuscript, and 96.1% hallucination capture rate on 10,000 production documents across five document types. The Data-Driven Query Generator addresses NL-to-SQL at its root cause by inverting the generation direction, pre-computing the inverse mapping from SQL to natural language from real data and validating every pair by actual database execution. The method is validated across four heterogeneous SQL engines on 6,000 production pairs with cross-engine execution error below 0.3% and 94.2% execution accuracy at sub-millisecond latency. Both methods are integrated in MIKA (Multi-modal Intelligent Knowledge Analysis), where three-level distributed parallelism over atomic operations yields 69.9× speedup, making hallucination detection viable at production scale — 2,644 pages/min on three bare-metal nodes, processing up to 5 million documents per day across healthcare, banking, and fiscal domains.
Jaime et al. (Sun,) studied this question.