LLMs use information at the beginning and end of their context better than the middle.

LLMs use information at the beginning and end of their context better than the middle. Mitigation: place the most important retrieved chunks at start/end of context. Use fewer, higher-quality chunks (reranking helps). Ask the model to cite specific sections, forcing it to attend to the full context. Long-context models reduce but don't eliminate the problem.