Is the model actually writing tool calls to show results on the map? Pretty cool idea. There are some minor visual bugs on the demo you might want to patch though.

Demo is just a replay of old conversations, but if you're doing your own chats then they are streamed to the map as the llm mentions them. Experimented a lot with tool vs no tool, landed on instructing the model to embed XML tags with the response and parsing them while streaming, since then you don't have to break the LLM response process with tool calls.

In my experience this is faster and cheaper (at least for Claude, since caching costs there) compared to tool calls for registering map results.

Thanks for the feedback on the demo, which browser are you using so I can check that?