What are your two cheapest levers against lost-in-the-m...

To address the lost-in-the-middle phenomenon without retraining, you should use a combination of prompting and structural reordering. A prompt-based fix involves explicitly instructing the model to pay attention to the middle of the context window. A structural fix involves reordering the retrieved candidates so that the most relevant information is placed in the model’s ‘safe zones’ (the beginning and end of the context). Reordering is considered more reliable because it changes the input distribution to align with the model’s natural performance characteristics, whereas prompting asks the model to actively fight against its own trained positional biases.