I tried asking the assistant for court rulings on a complex subject, asking for the relevant extracts. I tried a lot of variations of the prompt to ensure that the extracts were not hallucinated. None of the models I tried gave me the right answer, and all citations were hallucinated.
I have seen this happen quite a lot when using the assistant for legal questions.
Could there not be a way to ensure that citations in the assistant output actually are in the source document ? The system prompt could be configured to enclose citations in some markup that can be parsed and the citation checked for actual presence in the source ? A citation tool ? A combination of both ?
The assistant would not wrongly quote sources. Also, a link with a highlight would be great.