Onton just dropped a search model that doesn’t play by the usual rules. Ontology 1, their new neurosymbolic system, scored a 0.630 mean precision@10 on a 90-query home decor benchmark, leaving Google Shopping at 0.543 and Amazon at 0.469 in the dust. It pulled this off while indexing roughly 1% of either giant’s product catalog. Three independent LLM judges—Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5—scored the results, and while inter-judge agreement was modest (Krippendorff’s alpha of 0.465), the rankings were unanimous: Onton on top, then Google, then Amazon.
The magic is in the architecture. Conventional e-commerce search treats intent as a filterable set of attributes: size, price, brand. That falls apart when a shopper types “pet-friendly sectional.” There’s no filter for that, and seller labels are frequently missing or outright lies. Ontology 1 doesn’t trust the listing. It reasons from objective properties—fiber type, weave, construction—and builds an explicit world model. When it encounters an unfamiliar concept like “pet-friendly,” it works backward: durability and cleanability matter, so polyester upholstery becomes an indicator. That reasoning gets reused for “pet-friendly chair” or “cleanable blue couch,” creating a continuous learning loop.
This isn’t a model you can download. Ontology 1 is live at Onton.com, and partner access is granted case-by-case for teams building agentic-commerce platforms. There’s no public API, pricing tier, or open checkpoint. For now, the index covers only home decor and furniture, which makes the narrow catalog breadth a genuine weakness. The model stumbled on queries like “lamp that won’t wake my partner if I read at 3am” (Onton 0.4 vs. Amazon 0.9) and “something to put on a weirdly deep windowsill” (Onton 0.07 vs. Amazon 0.67). Amazon’s massive category metadata won those rounds.
Under the hood, Ontology 1 runs on Ograph, a custom graph database that Onton claims crushes SuiteSparse:GraphBLAS—one Ograph core delivers roughly 100x the throughput per core, with a GPU build running 43x faster than the CPU variant. Early tuning suggests that could hit 1000x. This kind of infrastructure matters because the whole system depends on building and querying an inspectable knowledge graph in real time, not just retrieving vectors from an index. The real question is whether this approach scales beyond a single vertical. Onton says the methodology generalizes to non-product data with essentially no reconfiguration, but for now, they’re proving it one sofa at a time.
