Tiny fixes

[anthony2261]

• return toolmessage with error if LLM calls QuerySQLDataBaseTool without for_chart
• use gpt-4o-mini for title generation

[RamiAwar]

Handling a bad tool call such as this seems like it should be the responsibility of whatever LLM calling framework we're using, not good to start handling these manually.

Maybe we should look at what's being used here and fix the issue at the higher level automatically (by enforcing a structure for the sqldb tool call I mean). Makes sense?