A retrieval system makes several decisions before a reader sees a result: how to represent material, which candidates to retrieve, how to order them, and how much work to perform along the way. A gain at one step can conceal a loss at another. More relevant candidates in the first hundred do not necessarily give the reader a better first ten.
These essays follow those decisions in order. Start with a geometric example of shared references, examine a distinction those references can lose, then turn to our retrieval diagnostic and a concrete performance repair. The mathematical examples, historical evaluation, and engineering measurements support different kinds of conclusions; each essay explains what its result can tell us.
Connecting vector spaces through shared references ↗
Understand why equal vector length does not create a shared coordinate system, and try the rotation example.
What shared references leave out ↗
See exactly how a chosen reference set can collapse two distinguishable objects into one representation.
For now, we are keeping BM25 ↗
Read the recall and top-ten results separately before deciding whether a retrieval blend deserves to replace BM25.
A query visited 11 agents—and paid for 512 ↗
Trace a local query that paid for full-network preparation, then inspect the same workload after the repair.
To use this reading path on your own system, choose one request and follow it all the way through. Record what information each representation preserves, whether each stage changes membership or only order, and which preparation costs the end-to-end measurement includes.