Prototype: Grounding AI and LLMs with Overture's Cross-Theme Knowledge Graph

Overture Maps Foundation is testing whether map geometry can serve as a connective layer across data themes, with a prototype called ORATOR built by member Wherobots generating 700,000 nodes and 1.2 million edges over the San Francisco Bay Area to create a cross-theme knowledge graph for grounding AI and LLMs. The prototype uses spatial predicates to automatically link entities across themes, aiming to eliminate the 'conflation tax' that developers currently pay to manually join spatial data.

LLMs are good with text and bad with coordinates, so they guess when asked where things are. Overture is testing whether map geometry itself can become the connective layer across data themes, giving AI a verifiable spatial graph to reason over. ORATOR, a prototype built by Overture member Wherobots https://wherobots.com/ , led by Daniel Smith https://www.linkedin.com/in/daniel-smith-spatial/ generated 700,000 nodes and 1.2 million edges over the San Francisco Bay Area to see if the idea holds. Geometry is the foreign key. If a place sits inside a building, the data already knows they are related. Nobody had to write that link by hand. This effort is in the prototyping phase, and we want input from the Overture community before any of it hardens into a standard. The goal is a cross-theme knowledge graph that solves real problems for developers and AI engineers, not a clever demo. A strategic pillar for AI readiness Large language models process language well. Ask one to do raw coordinate math or reason about what is near what, and it starts to struggle. An AI agent routing a delivery or sizing up the businesses on a single city block needs more than a good vocabulary. It needs a structured, verifiable grounding layer https://overturemaps.org/blog/2026/open-spatial-location-grounding-for-ai/ it can trust. Positioning Overture’s open map data and the Global Entity Reference System GERS https://docs.overturemaps.org/gers/ for clean AI and LLM consumption is a core strategic priority for the foundation. A trusted source of geospatial facts and relationships is what stops a model from inventing spatial context when it does not actually know the answer. The invisible conflation tax Overture assembles high-quality data from many open sources across Places, Buildings, Addresses, Transportation, and Divisions. Open data arrives this way by nature: independent feeds, built by different communities for different purposes, with no shared keys linking a place to the building it occupies or a building to the road that serves it. Within a single theme, Overture already provides these links – Transportation and Divisions carry their own connective structure. The open question is how to link across themes, where the same real-world entity may show up as a point in one feed, a polygon in another, and a line in a third. So a plain human question, like “which businesses operate inside this building footprint,” forces developers to write their own spatial joins before an LLM can touch the data. That work gets repeated by every team that needs the same answer. We call this the conflation tax: a recurring cost the whole downstream ecosystem pays to rebuild relationships that the geometry already implies. Every AI team grounding a model in the physical world ends up solving the same join problem. The conflation tax is what they pay to do it alone. Weaving the graph: a conceptual prototype To explore solutions, the Overture Product Council has been evaluating approaches to cross-theme linkage. As part of that work, Overture member Wherobots built an early-stage prototype called ORATOR Overture Maps Foundation Knowledge Graph . This is an experimental proof-of-concept, not a finished at-scale product. It is a testing ground for one question: how should spatial relationships be structured to serve AI workflows well? Wherobots has written up the broader technical pattern behind this work in Graph RAG for the Physical World https://wherobots.com/blog/spatial-graph-rag/ , which walks through the spatial joins, storage, and graph traversal in detail. ORATOR runs on pure spatial derivation. The working principle is that geometry acts as the foreign key. When a place point falls inside or near a building polygon, a spatial predicate turns that physical fact into a graph edge automatically. Key features under test: - Relationships carry confidence scores. Strict containment earns a 1.0; weaker spatial signals score lower. - Fallback rules add hierarchy. An address inside the same building as a place beats an address that is merely closer in raw distance. - Every edge keeps full provenance, so a downstream model can trace exactly why and how two entities were connected. Under the hood: testing the architecture The prototype shows how separate themes might become a working spatial knowledge graph. - Six node types are derived from Overture’s core themes: building, place, address, connector, division, and a derived snap point for road access. - Nodes connect through eight semantic relationship types, including located in, has address, and access. - An experimental MCP server exposes the graph to AI agents through query tools such as extract subgraph and find nearby, so a model receives structured knowledge rather than a pile of coordinates. - In a San Francisco Bay Area proof of concept, the pipeline produced more than 700K total nodes and 1.2M total edges in roughly 8 minutes. The same pipeline was applied to an AOI over Manhattan and produced ~400k nodes and 3.5m total edges in roughly 10 minutes. - Access edges use an inheritance model. Giving every address, place, and building its own road-access edge would mean more than 600,000 edges. Instead, a place inherits the access edge of the building that contains it, which collapses that naive count to roughly 172,000 semantically correct access edges. - Counts, node and edge taxonomies, and build time are drawn from the ORATOR prototype run summarized in the methods note below. A model should not have to re-derive the map every time it answers a question. Pre-computed edges turn “probably related” into “structurally coherent.” Call for feedback: help shape the knowledge layer The Overture Product Council meets bi-weekly to discuss the product direction of Overture. Join Overture to attend these meetings and help provide feedback on these three open questions: Schema versus downstream product . Should cross-theme relational links live in the core Overture schema, or are they better delivered as a specialized downstream product? Prioritization . Which cross-theme edge types, for example located in or has address, deliver the highest immediate return for LLM grounding? Metadata standards . How should confidence scores and provenance be standardized so that LLMs can weigh the truth of a relationship appropriately? A standardized, cross-theme spatial knowledge layer could support a wide range of enterprise work, from logistics and asset management to last-mile delivery and AI grounding. Ready to ground your AI? - Explore Overture’s vision for open spatial and location grounding https://overturemaps.org/blog/2026/open-spatial-location-grounding-for-ai/ . - Join Overture to shape the future of the knowledge layer. Visit the Overture website http://google.com/url?q=https://overturemaps.org/&sa=D&source=docs&ust=1783520341530908&usg=AOvVaw3cqm7LON3BaN50R8vyLETV to learn how your organization can become a member. Explore Overture’s reference data and system https://docs.overturemaps.org/ through the developer documentation.- Read Wherobots’ companion write-up, Graph RAG for the Physical World https://wherobots.com/blog/spatial-graph-rag/ , for the technical pattern behind the prototype. - For the latest schema releases and AI tooling, sign up for the monthly newsletter https://share.hsforms.com/1T38OzgVmQPG7AQTdAyBizQ4tvhy? hstc=180738424.c4ad53cb574a5cc23150c94b83f9e553.1746112778982.1762779637776.1762785419816.36& hssc=180738424.1.1762785419816& hsfp=548571637 or follow Overture on LinkedIn https://www.linkedin.com/showcase/overture-maps-foundation/ , X https://x.com/OvertureMaps , and Bluesky https://bsky.app/profile/overturemaps.bsky.social . Methods and data provenance The quantitative claims in this post come from a single ORATOR prototype run over the San Francisco Bay Area test region, executed on Wherobots Cloud. The figures below are from that run and are directional proof-of-concept results, not benchmarked production numbers. Metric | San Francisco Bay Area | Manhattan | Source | |---|---|---|---| | Total nodes | 700,000+ | 400,000+ | ORATOR run | | Total edges | 1.2M+ | 3.5M+ | ORATOR run | | Build time | ~8 min | ~10 min | ORATOR run | | Buildings | 170,965 | 58,379 | Overture theme | | Places | 55,134 | 159,463 | Overture theme | | Addresses | 387,872 | 76,794 | Overture theme | | Road segments | 80,931 | 46,032 | Overture theme | | Road connectors | 83,765 | 71,621 | Overture theme | | Divisions | 102 | 92 | Overture theme | | Division areas | 43 | 49 | Overture theme | | Naive access edges one per address + place + building | ~610,000 potential | 294,636 | Hypothetical | | Access edges after inheritance | 172,053 | 59,323 | Derived inheritance model | | has address edges | 437,757 | 234,793 | Derived two-pass linking | Method notes for reviewers: - Distances use geometry-to-geometry computation rather than centroid-to-centroid for polygons and linestrings useSpheroid=false with reprojection to EPSG:4087 . - The naive access count of roughly 610,000 is the sum of all addresses, places, and buildings 387,542 + 52,165 + 170,865 under a scheme where each entity gets its own road-access edge. - That scheme was never built – the inheritance model a place inherits its building’s road access produces 172,053 access edges instead. - Those 172,053 access edges, plus 437,757 has address edges, plus the road network and the remaining spatial edges, sum to the 1.2 million total. - Address linking runs two passes: containment and street-name match first at confidence 1.0, then a proximity fallback scored 0.6 to 0.95. - The spatial joins, Iceberg storage, and graph traversal behind these figures are documented publicly in Wherobots’ Graph RAG for the Physical World.