Diagnosing RAG Failures: A Taxonomy of Failure Modes in Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) has become a standard paradigm for grounding large language models in external knowledge, yet RAG systems still fail in production in ways that are poorly understood and rarely standardized. Existing work focuses on architectures and accuracy benchmarks, leaving practitioners without a common language for describing and diagnosing RAG failures. In this paper, we propose a failure-centric taxonomy of five modes in RAG systems: Retrieval Blindness, Retrieval Noise, Evidence Ignorance, Evidence Over-Reliance, and Latency-Induced Failure. For each mode, we define typical patterns and causes and illustrate them with concrete scenarios. We further introduce a practical diagnosis guide that maps observable symptoms to failure modes and candidate mitigations, and illustrate the taxonomy through four enterprise case study scenarios. Building on this conceptual framework, we contribute a preliminary empirical validation of three of the five modes---Retrieval Blindness, Retrieval Noise, and Evidence Ignorance---using two fixed generator models (Claude Sonnet 4.6 and Llama 3.3 70B Instruct) under an identical retrieval pipeline, prompt template, and evaluation protocol, with controlled clean and stress conditions drawn from the RGB and Lost-in-the-Middle public benchmarks. We report token-level F1, retrieval recall/precision/nDCG, exact match, cost, and latency across 900 real API calls (450 per model), together with a manual error-categorization pass on a stratified sample of 70 outputs. Both models show substantial, consistent degradation from clean to stress conditions across all three tested modes, supporting the taxonomy's claims for these three modes only---Retrieval Blindness, Retrieval Noise, and Evidence Ignorance---for these two models on this dataset; the remaining two modes (Evidence Over-Reliance, Latency-Induced Failure), a third model (GPT-4.1), and broader taxonomy-wide validation all remain future work. We are explicit that this is a small-scale, two-model, single-dataset-family preliminary study, not a full validation of the taxonomy, and we detail its limitations alongside the results. Our aim is to establish both a shared vocabulary/diagnostic tool for RAG failures and a reproducible empirical basis that future work can extend.