Onton’s new Ontology 1 model outperforms Google Shopping and Amazon on a multimodal product‑search benchmark, using a neurosymbolic knowledge graph to reason about vague queries.

Onton has unveiled Ontology 1, a neurosymbolic search model that claims a 2.7× accuracy boost over leading e‑commerce search engines such as Google Shopping and Amazon on a multimodal product‑search benchmark.

What Is Ontology 1?

Ontology 1 combines deep‑learning vision and language encoders with a symbolic knowledge graph that captures product attributes, relationships, and taxonomies. The hybrid architecture lets the system reason about vague or underspecified queries, translating them into precise attribute constraints before retrieving results.

Benchmark Performance

In Onton’s internal benchmark, which mixes images, text descriptions, and user‑generated queries, Ontology 1 achieved a mean reciprocal rank (MRR) 2.7 times higher than the best publicly reported figures for Google Shopping and Amazon’s search stacks.

The test set includes over 100,000 product listings across categories such as apparel, electronics, and home goods, and features queries ranging from exact product names to abstract phrases like “cozy winter outfit” or “budget-friendly laptop with long battery life.”

How Neurosymbolic Reasoning Improves Results

Traditional neural search pipelines rely on vector similarity, which can miss nuanced intent when the query lacks specific keywords. Ontology 1 first maps the query onto the knowledge graph, extracting latent concepts (e.g., “cozy” → material softness, “budget‑friendly” → price range). It then re‑weights candidate products based on graph‑derived constraints, producing a shortlist that better matches user expectations.

  • Semantic grounding of vague adjectives
  • Dynamic attribute inference from related items
  • Cross‑modal alignment of image and text features
  • Explainable ranking scores derived from graph paths

Potential Impact on E‑commerce

If adopted broadly, Ontology 1 could reduce the “search‑to‑purchase” friction that plagues online retailers, especially for shoppers who rely on natural language or visual cues rather than exact SKU numbers.

The model also opens the door to more transparent recommendation systems, as the symbolic layer can be queried to explain why a product was ranked highly—something pure neural models struggle to provide.

Onton plans to release an API later this year, allowing third‑party platforms to integrate the neurosymbolic engine without rebuilding their own knowledge graphs.

For a detailed look at Ontology 1’s architecture and benchmark methodology, see the MarkTechPost coverage of Onton’s Ontology 1 launch.