Hmm, setting the "chat_mode" explicitely to "condense_question" seems to fix the issue as this is supposed to allways query the query engine. Interestingly, leaving it unspecified, it fails to return the published date. Isn't "condensed_question" supposed to be the default mode?
Actually, the default mode is using an agent. Which, based on the description of the index, the LLM decides whether it queries the index or not.
The context and condense_question mode will use the index every time. Context usually works ok-ish, but it helps to set a system prompt to better instruct how the LLM should act (that could be said for all chat modes though)
Imo condense question is ok-ish, but I personally find it a little slow + some unnatural interactions.
