Guys, maybe this will be of interest to some of you. I’ve been building a LLM-driven pipeline to pick the right tag keys and values when querying local Geofabrik/PostgreSQL data, based on the separately parsed user query. The flow is two-stage: first, the LLM proposes the most discriminative key for the object type (from an allowed key list, which is all columns from Geofabrik tables). Then, instead of guessing values, I pull the real values that actually exist in the country’s precreated tag-value index, feed that list to the LLM, and let it choose the best matches. There’s a parallel “hstore” path where the model picks a single best hstore key from an allowed set, and a second step selects the matching value from real data. It’s been a big win for precision: the LLM never invents tags — it only selects from what’s truly in the database, so scans are faster and cleaner.
Are there any similar frontends for OSM, do you know? Mine is available for testing at www.locationvault.com.
