Architectural Hallucination Mitigation in Multi-Agent Document Intelligence: Domain-Isolated RAG and Dual-Channel Claim Verification

Hallucination in large language model (LLM) pipelines is the central unsolved problem in enterprise document intelligence: models produce fluent, structurally plausible outputs factually unsupported by the source document. Existing mitigation strategies operate within the generative process — retrieval augmentation, self-consistency, chain-of-thought grounding — and reduce but do not eliminate this failure mode. We address hallucination at the architecture level through three mechanisms implemented within MIKA (Multi-modal Intelligent Knowledge Analysis), a five-layer multi-agent platform for intelligent document processing. (C1) Domain-isolated multi-query RAG. Each of MIKA's 12 reasoning agents maintains a strictly isolated vector partition (Kᵢ ∩ Kⱼ = ∅) and issues four independent queries per analytical dimension, eliminating cross-domain retrieval contamination architecturally. A three-stage ablation on KGPPS clinical NLP validates C1: shared-index baseline MAD = 2.16; domain isolation with equivalent content MAD = 1.91; full C1 MAD = 1.58 — isolating a 0.33 causal contribution of partition isolation above content enrichment alone. (C2) Dual-channel claim verification. A deterministic source channel S accesses the source of record independently of the LLM; every structured claim is verified against Ts before delivery — claims without grounding are refused rather than hallucinated. Calibrated on 1,240 documents (8,431 claims), C2 reduces delivered hallucination from 10.0% to 0.29% at 97.1% claim precision. The pattern generalizes to any domain with a machine-accessible ground truth (OCR, ERP, HL7/FHIR, contract text). (C5) Schema-free, training-free extraction. Extraction targets are runtime natural language specifications validated by C2. On 180 documents spanning four heterogeneous types, C5 achieves 91.7% field accuracy vs. 73.8% for a template-based industrial baseline — without retraining.