Hey, not sure how much this will be helpful but in our case, we use a custom AI server that's hosted by us. We use Qdrant as our vector DB (to analyze text documents, generate summaries and reports, etc.) and a readonly (!) SQL adapter for querying databases. The AI agent is informed on the table structure and can execute its own SQL queries. It does this in a code execution environment where the query API is exposed, it also has access to pandas and scikit for data science stuff.
We handle financial reports and accounting and found that this works relatively well, still nowhere near what Julius AI offers though