Seven Recurrent RAG Failure Points

Empirical studies of operational RAG systems identify seven recurring failure modes:

1. Retrieval errors: wrong documents retrieved
2. Context consolidation failures: relevant information not properly combined
3. Hallucinated outputs: generation diverges from retrieved evidence
4. Incomplete answers: partial information used, leaving gaps
5. Misaligned evidence: semantically matching but factually irrelevant content
6. Redundancy overload: too much repetitive context degrading generation
7. Latency cascades: pipeline inefficiencies compounding under load

This taxonomy (Barnett et al., 2024) is useful for diagnosing production RAG issues. Most failures aren’t random, they cluster into recognizable patterns that suggest specific architectural interventions.

The meta-observation: RAG isn’t failing in novel ways. It’s failing in predictable ways that repeat across implementations.

Related: 05-atom—rag-core-equation, 05-molecule—rag-evaluation-dimensions, 05-molecule—rag-architecture-taxonomy