Onton, a San Francisco-based search and discovery company, has released Ontology 1, a neurosymbolic model for complex, conversational, multimodal product search. On a 90-query benchmark scored by three independent LLM judges, Ontology 1 reached a mean precision@10 of 0.630, against 0.543 for Google Shopping and 0.469 for Amazon. It did this while indexing roughly 1% of their catalogs.
Is it deployable
Yes, but not as weights you download. Ontology 1 is live for end users at Onton.com, and Onton says partner access is granted case by case for teams building on the agentic web. There is no public API, pricing tier, or open checkpoint for the model itself. Adoption today looks like a partnership, not a pip install.
- Company fit: Mid-market and enterprise retailers, marketplaces, and agentic-commerce platforms whose relevance stack already loses on long, requirements-heavy queries. Small catalogs see less benefit, because the failure mode Ontology 1 targets scales with catalog size and listing noise.
- Industries: Home decor and furniture today, since that is the only vertical Onton indexes. Onton states the methodology generalizes beyond e-commerce, and that Ontology searches non-product data with essentially no reconfiguration.
- Applications: Conversational and multimodal site search, moodboard-driven discovery, negation-heavy filtering, listing and review trust scoring, and grounding layers for shopping agents.
Why keyword and vector retrieval break here
Conventional e-commerce assumes intent maps onto categories and attributes: size, price, material, brand. There is no filter for ‘pet-friendly,’ and none for furniture that fits your room. Onton argues this catalog interface has barely changed in nearly 30 years.
Ontology 1 takes a different route. For ‘pet-friendly sectional,’ it does not trust the seller’s label, which may be absent or untrue. It reasons from properties more likely to be objective — fiber, weave, construction — and flags claims the product data contradicts. It also weighs the source, since some listings game the algorithm and some reviews are bought.
The model builds an explicit, inspectable world model rather than absorbing patterns into weights. When it has no account of ‘pet-friendly,’ it treats that as a gap and works the answer out: cleanability and durability, then polyester upholstery as an indicator. The learning is reused on later queries such as ‘pet-friendly chair’ or ‘cleanable blue couch,’ and the loop runs continuously.
The benchmark: Subtext-Decor-90
Onton released Subtext-Decor-90 with code and data. Three multimodal judges: Claude Opus 4.8, Gemini 3.1 Pro and GPT-5.5, scored the top 10 visible result cards returned by Onton, Amazon and Google Shopping for each of 90 text queries. P@10 was averaged across 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 won 52 queries outright, Google 19, Amazon 16. Those sum to 87 because Ontology returned fewer than 10 results on three queries, and empty slots were scored as non-relevant. Excluding those slots instead gives Onton 0.665, Google 0.549, Amazon 0.459.
Krippendorff’s alpha across the three judges is 0.465, so absolute P@10 values are noisy and judge-dependent. All three judges still place the engines in the same order.
Image and multimodal queries were excluded from the 90, because Amazon Lens does not support multimodal queries and Google Lens does not return products exclusively. Onton reports a separate 10-query image and multimodal comparison against Google.
Where Ontology 1 loses
Failure cases cluster on functional-spec queries where Amazon’s category metadata dominates: ‘lamp that won’t wake my partner if I read at 3am’ (Onton 0.4, Amazon 0.9) and ‘something to put 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 narrow the gap.
The infrastructure underneath
Ontology 1’s knowledge graph runs on Ograph, a custom graph database. Onton reports one Ograph core beating SuiteSparse:GraphBLAS running on 14 cores, roughly 100× the throughput per core, and a GPU build running 43× faster than the CPU variant, with early runs touching 1000× as the implementation is tuned.
Interactive explainer
The embed below walks through the same material in four panels: real 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.
