Source Reconstructibility for Robust Document Processing and Relational Querying with LLMs — Verification, Domain Partitioning, and Data-Driven Query Catalogs in Enterprise Settings

LLMs deployed in enterprise document processing and relational query systems exhibit three structurally distinct failure modes that current mitigation techniques cannot resolve: (1) value hallucination on precision entities in document extraction, where a single-digit error on a dosage, amount, or tax code carries direct legal or clinical consequence; (2) cross-domain contamination in retrieval-augmented generation, where retrieval over indices that mix domains produces plausible but ungrounded completions; and (3) ill-posed NL-to-SQL, where the combinatorial unboundedness of natural language makes any translation function structurally incorrect over an unbounded input domain. This paper presents three contributions that address each mode at its root, articulated by a common operational principle: preservation of source reconstructibility. VERA (Verified Extraction with Region Anchoring) provides reconstructibility over document extraction via the OCR ∘ LLM ∘ VERA composition, with multi-layer verification (region anchoring, inter-field coherence, fuzzy logic, and the Numeric Determinism Layer NDL) achieving 96.1% hallucination capture rate on 10,000 production documents. Domain Partitioning (PC-RAG) makes cross-domain contamination architecturally impossible under the constraint Ki ∩ Kj = ∅, validated against unified-index baselines. The Data-Driven Query Generator (DDQG) inverts the NL-to-SQL paradigm by replacing free user NL with selection from a pre-validated catalog built offline from metadata and real data, validated on 6,000 production pairs across four heterogeneous SQL engines (Oracle, SQL Server, PostgreSQL, MySQL) with cross-engine error below 0.3%. The three contributions are integrated in MIKA (Multi-modal Intelligent Knowledge Analysis), achieving 69.9× speedup and 2,644 pages/min throughput on three bare-metal nodes, with peaks of 5 million documents per day in commercial deployment. The paper additionally establishes a principle of technical verifiability for AI contributions in operational domains, aligned with the EU AI Act, DORA, and GDPR documentation requirements.