GlossaryEvaluation
Lost in the middle
The empirical finding that LLMs pay less attention to content placed in the middle of a long context window.
A well-documented phenomenon (Liu et al. 2023) showing that LLMs reliably attend to the beginning and end of their context but struggle with content placed in the middle. This implies that for RAG, you should put the most relevant retrieved passage either first or last - not buried among less-relevant ones. Good rerankers and tight chunking are how you exploit this.