|

Onton Releases Ontology 1: A Neurosymbolic Search Model That is 2.7x More Accurate than the World’s Best E-commerce Search Engines

ByRicardo

Onton, a San Francisco-based search and discovery firm, has launched Ontology 1, a neurosymbolic mannequin for advanced, conversational, multimodal product search. On a 90-query benchmark scored by three unbiased LLM judges, Ontology 1 reached a imply precision@10 of 0.630, towards 0.543 for Google Shopping and 0.469 for Amazon. It did this whereas indexing roughly 1% of their catalogs.

Is it deployable

Yes, however not as weights you obtain. Ontology 1 is stay for finish customers at Onton.com, and Onton says accomplice entry is granted case by case for groups constructing on the agentic internet. There is no public API, pricing tier, or open checkpoint for the mannequin itself. Adoption at this time appears like a partnership, not a pip set up.

  • Company match: Mid-market and enterprise retailers, marketplaces, and agentic-commerce platforms whose relevance stack already loses on lengthy, requirements-heavy queries. Small catalogs see much less profit, as a result of the failure mode Ontology 1 targets scales with catalog measurement and itemizing noise.
  • Industries: Home decor and furnishings at this time, since that is the solely vertical Onton indexes. Onton states the methodology generalizes past e-commerce, and that Ontology searches non-product knowledge with primarily no reconfiguration.
  • Applications: Conversational and multimodal web site search, moodboard-driven discovery, negation-heavy filtering, itemizing and overview belief scoring, and grounding layers for procuring brokers.

Why key phrase and vector retrieval break right here

Conventional e-commerce assumes intent maps onto classes and attributes: measurement, worth, materials, model. There is no filter for ‘pet-friendly,’ and none for furnishings that matches your room. Onton argues this catalog interface has barely modified in practically 30 years.

Ontology 1 takes a unique route. For ‘pet-friendly sectional,’ it doesn’t belief the vendor’s label, which can be absent or unfaithful. It causes from properties extra prone to be goal — fiber, weave, development — and flags claims the product knowledge contradicts. It additionally weighs the supply, since some listings recreation the algorithm and a few evaluations are purchased.

The mannequin builds an express, inspectable world mannequin reasonably than absorbing patterns into weights. When it has no account of ‘pet-friendly,’ it treats that as a niche and works the reply out: cleanability and sturdiness, then polyester upholstery as an indicator. The studying is reused on later queries equivalent to ‘pet-friendly chair’ or ‘washable blue sofa,’ and the loop runs repeatedly.

The benchmark: Subtext-Decor-90

Onton launched Subtext-Decor-90 with code and knowledge. Three multimodal judges: Claude Opus 4.8, Gemini 3.1 Pro and GPT-5.5, scored the high 10 seen outcome playing cards returned by Onton, Amazon and Google Shopping for every of 90 textual content queries. P@10 was averaged throughout judges, with 95% confidence intervals from 10,000 bootstrap resamples.

Results: Onton 0.630 [0.571, 0.688], Google Shopping 0.543 [0.490, 0.596], Amazon 0.469 [0.417, 0.521]. Onton received 52 queries outright, Google 19, Amazon 16. Those sum to 87 as a result of Ontology returned fewer than 10 outcomes on three queries, and empty slots had been scored as non-relevant. Excluding these slots as a substitute provides Onton 0.665, Google 0.549, Amazon 0.459.

Krippendorff’s alpha throughout the three judges is 0.465, so absolute P@10 values are noisy and judge-dependent. All three judges nonetheless place the engines in the identical order.

Image and multimodal queries had been excluded from the 90, as a result of Amazon Lens doesn’t help multimodal queries and Google Lens doesn’t return merchandise completely. Onton experiences a separate 10-query picture and multimodal comparability towards Google.

Where Ontology 1 loses

Failure circumstances cluster on functional-spec queries the place Amazon’s class metadata dominates: ‘lamp that received’t wake my accomplice if I learn at 3am’ (Onton 0.4, Amazon 0.9) and ‘one thing to placed on a weirdly deep windowsill’ (Onton 0.07, Amazon 0.67). Onton attributes this to catalog breadth and its single-vertical, non-sponsored index, and expects the self-learning loop to slender the hole.

The infrastructure beneath

Ontology 1’s data graph runs on Ograph, a customized graph database. Onton experiences one Ograph core beating SuiteSparse:GraphBLAS operating on 14 cores, roughly 100× the throughput per core, and a GPU construct operating 43× sooner than the CPU variant, with early runs touching 1000× as the implementation is tuned.

Interactive explainer

The embed beneath walks via the identical materials in 4 panels: actual Subtext-Decor-90 queries with per-query scores, the pet-friendly reasoning graph drawn step-by-step, the self-learning loop, and the benchmark chart with confidence intervals and alternate scoring views.