A language model asked for references will happily invent them. The failure is not fixable by prompting; it is structural, because the model is completing text rather than looking anything up.
AI Search in SciExpert inverts the order. A query first hits the open OpenAlex scholarly index and retrieves real records with DOIs, titles, venues and years. Only then is a model asked to synthesise an answer, and only from the retrieved set.
The practical consequence is that every claim in the synthesis maps to a record you can open. If a claim has no source behind it, it does not appear.
This is slower than pure generation and that is the point. The value of a research tool is not fluency, it is that you can hand the output to a reviewer without apologising for it.