Why would 200-token chunks with 20% overlap be wrong fo...

While fixed-size chunks work for simple L1 or L2 retrieval, they are inappropriate for compliance-based procedures. When you chunk a procedure into arbitrary 200-token segments, you often provide the model with only a subset of the necessary steps (e.g., steps 1-3 of 9). The model then executes these partial instructions confidently, leading to a wrong conclusion. The correct fix is to implement a secondary retrieval unit specifically for procedure-based content rather than adjusting the global chunk-size hyperparameter, which would negatively impact the majority of your traffic.