179 questions
No questions match those filters.
← All questions
Retrieval & EmbeddingsVery Hard· BM25· Evaluation· Corpus Analysis
A colleague retunes production to k1 = 0.9, b = 0.4 bec...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansThe improvement in recall@1000 may be misleading because it only measures if a document is in the pool, not its rank. In a corpus with a 400x length spread, the b parameter is critical for discounting long documents. If you lower b, you may be failing to properly discount long documents, which could degrade the top-10 results even if the overall recall improves. Request a per-length-decile breakdown of nDCG@10 before approving the change.