The retrieval opinion you hold that most people here would disagree with

[akash-coded]

One per reply. Argue for it in a paragraph. Bring a number if you have one — but an opinion
you cannot yet support is still worth posting, because someone may have the measurement.

To start, three of mine, all of which I have been argued with about:

1 · Most reranking is an expensive identity function. If the reranker's features overlap the
first stage's, it re-sorts by information the pipeline already used. This repo measured one making
retrieval worse at every k. The general rule is in The Reranker
Mirror.

2 · Growing the eval set beats almost every engineering option, and nobody does it. A day of
generator work moves more decisions than a month of tuning, because it changes what you can see.
Every argument about whether a 2-point delta is real is an argument about n.

3 · Chunking strategy is over-discussed and analyzer configuration is under-discussed. People
will spend a week on semantic chunking and never check whether their tokenizer splits
ERR_CONN_RESET into three terms that appear in every document. One of those is worth a week.

The rule for this thread

Disagree with the claim, not the person, and prefer evidence to seniority. A thread where the
loudest person wins is a thread nobody learns from.