Retrieval's bottleneck is moving from finding to selecting

I found 2 things that work for me. First, something I call density RAG: the agent calls it with multiple queries and gets back a semantic "heatmap", a table with chunks as rows, queries as columns and similarity scores in the cells. From this, the agent chooses which chunks to retrieve. Second, for the noise problem, passing the chunks through a decision model usually cuts it down. This is just my experience, I haven't benchmarked it yet.

OP here! Happy to jump in to answer any questions. :)