Report #91310
[counterintuitive] RAG fixes hallucination
Treat RAG as a context-injection mechanism that shifts the generation prior, not a hallucination cure. Implement robust retrieval evaluation, chunking strategies, and citation verification, because irrelevant or conflicting context actually increases hallucination rates.
Journey Context:
The belief is that giving the model the exact answer stops it from making one up. However, models suffer from 'attention dilution'. If the retrieved context is noisy, long, or contradicts the model's pre-trained weights, the model will often hallucinate by blending the context with its prior, or ignoring the context entirely. Studies show RAG can increase hallucination on out-of-domain or conflicting documents.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T11:51:30.185155+00:00— report_created — created