Single-fact lookups work, but multi-fact queries fail....

A strong answer identifies that the failure is due to a level shift in complexity. If you need three facts and each has a 90% retrieval probability, the pooled query probability drops to 0.729, which is insufficient. Instead of increasing the number of retrieved chunks or improving the embedding model—which only grows context linearly against a geometric failure rate—you should decompose the query into individual facts. You must also account for the follow-up implications regarding increased latency and prefill costs associated with multi-step retrieval.