subtopic

Compare methods and set a 'no answer' threshold

Fill the results table on the test set: BM25 vs embeddings vs embeddings + re-rank (Recall@1, Recall@3, MRR, time per query) and on your own 30 questions. Compare Recall@1 with Project 1's classifier accuracy — which would you deploy and when? Choose a score threshold below which the app says 'I could not find an answer — contact support'. Look at 20 failures.