How much should RAG retrieve?

Wix’s RAGXplain guidance suggests increasing retrieval depth when low context recall indicates that useful evidence is being missed. I tried that on WixQA, using its 6,221 Help Center articles and 200 expert-written questions. Each article stayed intact as one retrievable document, and I compared retrieval at k=1, 3, 5, 10, 20, 50.

It worked. Mean gold recall rose from 71.9% at k=5 to 83.2% at k=10 and 93.1% at k=20. After that the curve flattened: retrieving 30 more documents, from 20 to 50, added only 3.3 points of mean recall. Multi-article questions were harder. They continued to benefit from greater breadth because the first required article often ranked highly while a second required article appeared much later.

Multi-article questions stayed retrieval-hungry after the aggregate curve bent. At k=20, 69.2% had all required articles; at k=50, 82.7% did. In one case, a feature-request article saying something could not be done outranked the gold article explaining how to do it.

The practical lesson is small but useful: if retrieval is missing evidence, increasing k is worth trying before adding another mechanism. But more retrieval has diminishing returns. In this test, k=20 was where aggregate recall began to level off, while the additional results increasingly came from the same surrounding product neighborhood. That is not a recommendation to retrieve 20 documents everywhere. It is a reason to look at the curve on your own content before making the pipeline more complicated.