Poor retrieval quality is the primary driver of LLM hallucinations in RAG systems, not model size or prompt design. Research across HaluEval, TruthfulQA, and FaithDial benchmarks shows retrieval failures consistently dominate hallucination causes. Five key failure modes are identified: retrieval drift, context truncation, stale index poisoning, low-relevance top-k retrieval, and inter-agent miscommunication. Four dimensions of retrieval quality improvement are covered: embedding model selection, chunking architecture, retrieval strategy (hybrid search, cross-encoder re-ranking, relevance thresholding), and index freshness. In multi-agent pipelines, retrieval failures compound silently across agents, making detection especially difficult. Practical recommendations include auditing retrieval before upgrading models, implementing hybrid search, enforcing similarity thresholds, and validating context at every agent boundary.
