Onton Launches Ontology 1 Neurosymbolic Search Model

Onton, a San Francisco-based search and discovery company, has launched Ontology 1, a neurosymbolic model designed for complex, conversational, and multimodal product search. This new model demonstrated a mean precision@10 score of 0.630 across a 90-query benchmark, as evaluated by three independent large language model (LLM) judges. This performance significantly surpasses the 0.543 score achieved by Google Shopping and the 0.469 score by Amazon, despite Ontology 1 indexing approximately 1% of their respective catalogs. The model's effectiveness is particularly pronounced in handling long, requirements-heavy queries that traditional search methods struggle with. Ontology 1 is currently live for end-users on Onton.com, with partner access available on a case-by-case basis for teams developing on the agentic web. Onton has not released a public API, pricing tiers, or open checkpoints for the model, indicating that adoption currently involves partnerships rather than direct integration via a 'pip install.' The company positions Ontology 1 as a solution for mid-market and enterprise retailers, marketplaces, and agentic-commerce platforms whose existing relevance stacks falter on complex search criteria. Smaller catalogs are expected to benefit less, as the model's strengths lie in addressing the scaling challenges associated with larger catalogs and increased listing noise. Initially, Onton has indexed the home decor and furniture vertical, but the company asserts that the underlying methodology is generalizable beyond e-commerce to search non-product data with minimal reconfiguration. Potential applications include conversational and multimodal site search, moodboard-driven discovery, negation-heavy filtering, trust scoring for listings and reviews, and serving as grounding layers for shopping agents. Onton argues that conventional e-commerce search, which relies on mapping intent to categories and attributes like size, price, and material, has remained largely unchanged for nearly three decades. This approach fails to accommodate nuanced requirements such as 'pet-friendly' filters or furniture that precisely fits a specific room. Ontology 1 departs from this by not solely relying on seller-provided labels, which can be absent or inaccurate. Instead, it reasons from more objective product properties, such as fiber type, weave, and construction, to evaluate claims made in product data. This neurosymbolic approach combines symbolic reasoning with neural networks, enabling a deeper understanding of context and intent in search queries, particularly for subjective or complex product attributes. The model's ability to process and reason over unstructured or semi-structured data allows it to infer product suitability based on a wider range of characteristics than traditional keyword or vector-based retrieval systems.