Rethinking RAG Failures on Complex Queries: Multi-Intent Coverage and Constraint-Bound Validity Risk

Failures on complex queries in retrieval-augmented generation (RAG) are not uniform. Depending on query structure, they differ not only in retrieval form but also in user-facing meaning and risk. This study argues that complex queries in RAG should be differentiated at least into two types: Multi-Intent Queries (MIQ), in which multiple independent questions must be covered, and Constraint-Bound Queries (CBQ), in which a single question is governed by multiple jointly required conditions. Using a GOV.UK corpus, we conducted retrieval-only experiments on structurally differentiated query sets. The results show that MIQ is primarily a coverage / segmentation problem: performance improves substantially as chunk size increases, indicating that its main difficulty lies in whether both intended sources can be adequately covered. By contrast, CBQ shows only limited improvement, with top-ranked retrieval continuing to favor topically adjacent documents that satisfy only part of the required conditions. This suggests that CBQ is not merely a problem of retrieval granularity, but a structural problem of condition preservation and validity with stronger user-facing risk. These findings show that complex-query handling in RAG should not be evaluated only in terms of retrieval accuracy. Instead, failures should be understood in relation to query structure, user-facing risk, and the difference between incomplete retrieval and potentially misleading retrieval. This study provides a structural account of complex-query failure and offers an evaluative perspective for designing safer RAG systems.