01
Sources
OpenStreetMap, Open Data Euskadi and official GTFS with licence and date.
Back to the website
OPEN GEOGRAPHIC INTELLIGENCE INFRASTRUCTURE
emap/labs measures what AI knows about place.
Versioned urban datasets, bilingual geographic search, a production semantic service and a reproducible benchmark. The goal is not to answer every time: it is to answer with evidence and abstain when evidence is missing.
22geographic layers
27,515indexed places
139ES / EU cases
54strict held-out cases
ARCHITECTURE
Data before models.
Every result follows a traceable chain. Visual clients and agents reuse the same domain logic.
02
Catalogue
Reproducible pipelines, schemas, quality, coverage and freshness.
03
Hybrid retriever
High-precision keywords; e5-large only when semantics are needed; structured geo search.
04
Service
FastAPI on a VPS, explanations, reranking, rate limits and attribution.
05
Consumers
Web, mobile app, API and five MCP tools on one source of truth.
ES / EU BENCHMARK
Held-out decides.
The golden corpus evaluates hit@1, attributes, radius and abstention. The 54 held-out cases never select descriptions or thresholds.
| Retriever | dev ES | held-out ES | dev EU | held-out EU |
|---|---|---|---|---|
| baseline · keywords + geo | 74% | 60% | 71% | 62% |
| hybrid · MiniLM-L12 | 80% | 60% | 76% | 66% |
| hybrid · mpnet-base | 79% | 64% | 73% | 66% |
| hybrid · multilingual-e5-large | 76% | 73% | 71% | 71% |
Production · multilingual-e5-large · 1,024 dimensions · τ 0.80 · tie 0.01 · v2 evaluation dated 5 August 2026.
Measured abstention
19 answerable:false cases penalise a system that invents a category.
Validated Basque
Queries were checked with Itzuli; parity is measured, not assumed.
Two stages
Semantic category classification followed by geographic attribute and distance search.
Reproducible result
Every run records model, τ, tie-window, split and metrics in versioned JSON.
PRODUCTION
A small, explicit, measurable service.
The service loads layers in memory and exposes search, hike planning, accessibility, isochrones, freshness and data quality. e5-large runs on CPU through ONNX; keywords avoid semantic work when it is unnecessary.
92 mse5-large median
178 mse5-large p95
2.1 GBmodel on disk
57 msmeasured Jina reranker
DATASETS
Urban infrastructure, not personal profiles.
Release v0.2 contains 22 layers and 27,515 POIs. The hiking vertical crosses 2,825 peaks with 9 networks: 1,700 have a stop within 2 km as the crow flies; the real walking route is computed separately.
Fountains · toilets · AEDs · pharmacies
Bike parking · car parks · EV charging
Beaches · nature · peaks · shelters
Metro · Euskotren · rail · bus
Libraries · sport · accommodation
Neighbourhoods and infrastructure coverage
MCP FOR AGENTS
Five tools, narrow contracts.
The server is registered as io.github.r3tr0eth/emap and runs over stdio or Streamable HTTP. Every response keeps attribution and the do not pretend principle.
search_placesES / EU search with geographic context and abstention.
nearby_poisNearby services by layer, distance and source.
explain_placeMunicipality, neighbourhood and verifiable facts around a point.
plan_routeTransit, walk, bike and car routes through owned OSRM / OTP.
plan_hikePeak, access, stop, outward and return journey when verifiable.
ETHICS AND SCALE
Describe places without ranking people.
Labs has a hard rule: infrastructure, never people. It never ingests crime, income, demographics, origin or social proxies. Uneven OSM coverage is declared; neighbourhood indices identify service deficits rather than labelling places good or bad.