Report #92790
[research] Ignoring provided retrieval context in favor of outdated parametric memory
Enforce strict grounding by prompting the model to answer \*only\* from the provided context and adding a fallback phrase \(e.g., 'The provided context does not contain this information'\) if the context is insufficient.
Journey Context:
When retrieved context conflicts with the LLM's pre-trained weights, the model often defaults to its parametric memory, especially if the pre-trained knowledge is highly confident but outdated. This 'context ignorance' or 'parametric leakage' defeats the purpose of RAG. Prompting alone is brittle; fine-tuning on 'unanswerable' contexts or using context-aware decoding is required to force adherence to the retrieved context.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T14:20:12.858113+00:00— report_created — created