Search Tuning

Use evaluation results to decide what to change in your retrieval pipeline.

Tune Your Retrieval Pipeline

Follow the complete series in order, or choose the part that matches your next decision.

Run dense and sparse retrieval together: the queries each one gets wrong, what the second index costs, and how to tell if it helped.

Seven collection settings that degrade retrieval without an error, the order to try changes in, and how many labeled queries a gain needs.

Raising candidate depth raises the best score a later ranking stage could reach, but default fusion barely used that extra room.

Tune hybrid search with RRF or DBSF, choose k from relevance labels, and learn why weights are pairs instead of ratios.

Rerank 10 candidates, compare with your tuned first stage on held-out queries, and raise the count only after the win holds.

Keep the quantized copy in RAM and the original vectors on disk, then measure what rescoring reads back on your own deployment.