Exercises: Diagnose Chunking Failures
Use starter.py to classify failure traces and recommend controlled fixes.
1 Label the failure
Map symptoms to labels such as lost context, overlap noise, stale chunk, tiny chunk, giant chunk, or mixed topic.
2 Recommend a fix
Choose one fix candidate and avoid changing multiple variables at the same time.
3 Pick the metric
Select the metric that proves the fix worked: Recall@k, gold rank, duplicate rate, stale-hit rate, or citation support.
4 Write a production note
Return a compact failure-review summary that a team can use in a release discussion.