My understanding is that your levers are roughly better / more diverse embeddings or computing more embeddings (embed chunks / groups / etc) + aggregating more

My understanding is that your levers are roughly better / more diverse embeddings or computing more embeddings (embed chunks / groups / etc) + aggregating more cosine similarities / scores. More flops = better search w/ steep diminishing returns Colbert being a good google-able application of utilizing more embeddings. Search ends up often being a funnel of techniques. Cheap and high recall for phase 1 and ratchet up the flops and precision in subsequent passes on the previous result set.