Integration with popgetter geospatial agent

[sgreenbury]

This issue is for capturing integration with a geospatial agent (see Urban-Analytics-Technology-Platform/geollm#2 and Urban-Analytics-Technology-Platform/geollm#4).

An idea for the types of query we could aim to answer with an agent could involve retrieving population measures for a given place. The query could be broken down into:

• Identifying the place (e.g. a given city or region) and then using a tool to retrieve a bounding box for the region
• Identifying the type of metric and then returning either a suggested derived metric (fewer of these but might be simpler to identify a result) or a metric from the full catalog (many more metrics but might make mapping the query more challenging)

The final result returned could be a set of search params or a data request spec that can be used to preview and download the data.

[sgreenbury]

Another idea from @stuartlynn using embeddings for e.g. a query vehicle owners in Manchester

• Generate set of embeddings for all metrics once
• For each query, generate embedding for query and measure similarity to each metric
• Pick top few and return recipe to get that metric

[sgreenbury]

Initial version of this now implemented in popgetter (91-llm) and can be called from the CLI with a Qdrant running locally.

Next steps include:

• Rerunning for entire all metadata fields for a given metric (combined into e.g. JSON format)
• Containerise Qdrant and enable search functionality from the browser (e.g. as a new tab?) (see Add terminology and topics #21 for additional motivation for this).