What I would do first

If I inherited a retrieval system with no measurement, I would do these in order. Build a golden query set of a few hundred queries with graded relevance, including no-answer and version-sensitive cases. Add retrieval tracing so every production query is reconstructible. Measure recall at the context size, grounding rate at the claim level, and freshness compliance by content class, and publish the numbers. Then, and only then, start changing things: hybrid retrieval and reranking first, metadata filtering second, effective-dated retrieval with supersession third, progressive retrieval fourth. Each change measured against the set, each regression treated as a defect.

Retrieval quality is not a property of a model or a database. It is a practice. The teams that treat it as one end up with systems that get better every month, because they can see what is wrong. The teams that do not end up with systems that got worse at some point and nobody noticed.