popgetter-llm (experimental)

popgetter-llm provides library functionality for LLM applications with Popgetter.

Please refer to popgetter-cli (LLM integration) for a guide on using the LLM integration from the Popgetter CLI.

Aim and general approach

The popgetter-llm crate aims to build a natural language interface for Popgetter that enables the user to execute queries like:

"Produce me a file containing the population breakdown by age of men living in Manchester as a geojson"

An overview of the general approach take towards this is:

• Create some tools that interface with the popgetter library to execute generated recipes
• Create an LLM agent to generate those recipes from the query
• Allow that agent to be able to search our metadata for relevant metrics, geometries, etc for the query

Workflow

The LLM integration comprises two distinct parts: initialization and querying.

Initialization

• Populate the vector database with the embeddings for each metric description.

Query, search results and recipe

• User enters prompt for the LLM to consider the user's query and produce a list of datasets that will be required to answer it
• Generate an embedding vector from the prompt using the embedding API
• Find the top n closest vectors in the database
• Populate a new query with the metadata from those results along with the original query and prompt the generation of the recipe
• Run returned recipe to generate the resulting data