A hospital wants clinical notes to ground a chart-summary agent, but some scans still show another patient's sticker in the margin. What should the design do with those notes before grounding?
Select an answer to reveal the explanation.
Short Explanation
You would not drop a chart into the binder while someone else's name is still stuck on the corner. You filter those notes out before they ground the model, so the summary cannot see the other patient. A longer instruction does not peel the sticker off the page.
Full Explanation
Grounding text has to be clean before a model is allowed to read it. A note that still carries another patient's sticker is the wrong patient's data sitting inside this chart, and the design should filter that note out of the grounding set rather than hoping retrieval will be careful. Indexing every scan unchanged hands the mix to the retriever, and the model can quote the sticker as if it belonged on the chart. A longer system prompt is not a filter. The sticker remains in the retrieved text, and retrieved text is data the model can follow. Fine-tuning on the mixed stickers teaches the bad pattern and can make both names look like acceptable output. The filter belongs in preparation, before chunking and indexing, and it should quarantine the whole note when a second patient identifier is present. Leaving the page in the index keeps the wrong chart available to retrieval no matter how the prompt is worded. Exam caveat: this filter is a preparation step, and it is not a reason to train a custom model. Operational check: run a sample through the filter, confirm any page with a second patient identifier is excluded, and confirm a clean note for the right patient still enters the grounding set.