Domain-Partitioned Retrieval as a Hallucination Mitigation Strategy in Conversational RAG: The PC-RAG Architecture

Hallucination in Retrieval-Augmented Generation (RAG) systems is commonly attributed to generative model limitations. We argue that the primary cause in enterprise document retrieval is context window contamination: single-pass retrieval aggregates chunks from semantically heterogeneous domains, forcing the synthesis model to reason over a noisy cross-domain context and producing coherent but factually unstable responses. We present PC-RAG (Pipeline-Conversational RAG), an architecture built on domain-partitioned retrieval: complex queries are decomposed into domain-isolated sub-tasks, each serviced by an independent retrieval pass. Within each partition, intra-domain multi-query expansion generates k=6 semantic variants, maximizing coverage without cross-domain interference. Evaluation on a 1,200-document production deployment (HR, legal, compliance) shows that domain partitioning alone reduces false-positive chunk inclusion by 51% and raises human-rated synthesis quality from 3.1 to 4.1 on a 5-point scale (two annotators, κ=0.79, Wilcoxon p<0.001). The full PC-RAG system achieves a synthesis score of 4.2, outperforming HyDE Gao et al. [2023] (3.6), standard multi-query retrieval (3.7), and RAPTOR Sarthi et al. [2024] (3.9). RAGAS automated evaluation corroborates human judgements (r=0.83, p<0.001). By contrast, upgrading the synthesis model alone yields only marginal gains (3.1 to 3.4, p=0.031), confirming the bottleneck is retrieval architecture, not model capacity. The pipeline further integrates conversational coreference resolution (97.4% accuracy) and a semantic reranker; median end-to-end latency is 90 s, with primary contributions accounting for only 12% of this cost.