Generate interactive map layers from geospatial vector data?

[sfkeller]

You mention heat maps.

Would this approach allow interactive thematic map layers to be created that visualise geospatial vector data, such as that from OpenStreetMap (OSM)?

The natural language text input could be 'Map fountains and drinking water facilities in Lucerne'. The LLM would then extract the location (Lucerne) and named entities (fountains, drinking water amenity).

Possible MCP tools would be:

• get_location: geocoding geonames and retrieving map boundaries.
• get_tags: semantially match entities with appropriate OSM tags.
• get_query: do an NL2SQL and retrieve geospatial vector data from database (see PostPass).
• get_icons: match semantically entities to closest icon symbols.

[dippatel1994]

@sfkeller Thanks for the interesting idea and the Postpass reference!

Currently PaperBanana is focused on generating academic-style visuals from text, and the near-term plan is to expand domain coverage (finance, biology, etc.) while keeping the core task the same: text → polished visual output.

What you're describing feels like a different problem architecturally. More of a geospatial NL2SQL + map rendering pipeline than a visual generation task. Curious to hear your take: do you see this as a natural extension of PaperBanana's pipeline, or more as a standalone project that borrows the agent orchestration pattern?

Either way, I do have hands-on experience building text-to-SQL systems over large structured databases in production, so if there's interest in collaborating on this as a separate effort, I'd be open to exploring that.