context-degradation

Lost-in-middle detection maps attention weight across context positions and flags critical information sitting in the degraded middle region. The detect_lost_in_middle function compares critical token positions against the attention distribution and returns at-risk items with repositioning recommendations.

Context poisoning occurs when a hallucination, tool error, or incorrect retrieved fact enters context and compounds through repeated self-reference. Recovery requires truncating to before the poisoning point or restarting with verified-only context, since layering corrections over poisoned content rarely works.

Degradation typically begins at 60-70% of the advertised context window for complex retrieval tasks, with meaningful decline often starting at 8K-16K tokens. The curve is non-linear with a cliff edge, so benchmark your specific workload rather than trusting published thresholds.

Instructions placed in the middle of long contexts suffer 10-40% reduced recall due to the U-shaped attention curve. Move critical instructions to the beginning or end of the prompt, add explicit section headers as attention anchors, and trigger compaction before utilization exceeds 70%.

Contradictory retrieved documents cause context clash, where the model silently picks one source without signaling the conflict. Implement contradiction detection in the retrieval layer, establish source priority rules, and filter outdated versions before documents enter context.

Split tasks when a single context must maintain quality across dozens of independent objectives, since this creates degradation that more context cannot solve. Isolation is the most aggressive mitigation strategy but often the most effective for complex multi-task systems.