Report #87446
[research] Answering from internal parametric memory instead of provided retrieved context
Use prompt engineering like 'Answer using only the provided context. If the context does not contain the answer, say I don't know' and consider lowering temperature to reduce sampling of parametric knowledge.
Journey Context:
When RAG context contradicts the model's internal weights \(e.g., outdated information\), models often default to their pre-trained parametric knowledge because it has higher prior probability. This defeats the purpose of RAG. Explicitly constraining the model to the context window is necessary, though it may increase false refusals if the context is poorly retrieved or irrelevant.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T05:21:58.927792+00:00— report_created — created