# Ontology 1: A Neurosymbolic Leap for Conversational E-Commerce Search
Onton, a search and discovery company based in San Francisco, has unveiled **Ontology 1**, a neurosymbolic model designed to tackle complex, conversational, and multimodal product search. Designed to go beyond keyword matching, Ontology 1 builds an explicit, inspectable understanding of user intent to deliver superior results in e-commerce.
## Performance Validates the Approach
In a benchmark named **Subtext-Decor-90**, Ontology 1 achieved a **mean Precision@10 of 0.630**, outperforming Google Shopping (0.543) and Amazon (0.469). This test was conducted using three independent LLM judges scoring 90 text-based product queries across the top 10 results returned by each engine.
One of the key strengths of Ontology is its efficiency. It achieves these high scores while indexing only **roughly 1% of the retailers’ catalogs**, demonstrating that its neurosymbolic architecture is highly effective at reasoning over sparse data rather than relying on brute-force pattern recognition.
## Addressing the Core Limitations of Keyword and Vector Search
The limitations of traditional search are well-known. A keyword search is brittle, failing to understand synonyms or context. While vector (semantic) search improves relevance by understanding language, it can still be misled by poorly written listings or marketing spin.
Ontology 1 takes a different path. For a query like “pet-friendly sectional,” the model does not simply match text. Instead, it:
* **Identifies gaps in knowledge**: It understands that it lacks a direct filter for “pet-friendly”.
* **Decomposes the concept**: It reasons that pet-friendliness implies easy **cleanability** and **durability**, and that **polyester upholstery** is a strong indicator of these properties.
* **Weighs sources**: It accounts for noisy data and listings that are designed to game search algorithms.
* **Learns and reuses**: This reasoning process is cached and reused for future, similar queries, creating a compounding accuracy advantage.
## The Benchmark: A Closer Look at Subtext-Decor-90
To rigorously test its claims, Onton released **Subtext-Decor-90**, a benchmark dataset of 90 text queries covering complex decor and product scenarios. Three multimodal judges—Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5—independently scored the top 10 results returned by Onton, Amazon, and Google Shopping, using Precision@10 as the metric.
**Results (Aggregate Scores):**
* **Onton (Ontology 1): 0.630**
* **Google Shopping: 0.543**
* **Amazon: 0.469**
Onton outperformed its competitors in **52 out of 90 queries**, Google won 19, and Amazon won 16. This result was consistent even after excluding edge cases with fewer than 10 results.
## Where Ontology 1 Struggles (and How it Plans to Improve)
Ontology 1’s current limitations are well-defined. It performs less effectively on functional specification queries where Amazon’s vast, structured metadata provides a significant advantage. For example:
* *Query:* “lamp that won’t wake my partner if I read at 3am”
* *Performance:* **Onton 0.4**, **Amazon 0.9**
These “weaknesses” are viewed as a temporary state, as Onton’s continuous self-learning loop is designed to narrow this gap over time by integrating more data and feedback.
## The Infrastructure Powering the Intelligence
The model’s reasoning capability is supported by a robust technical foundation. Ontology 1’s knowledge graph runs on **Ograph**, a custom graph database. Initial benchmarks suggest a single Ograph core can outperform a traditional CPU-based graph processor by roughly **100x**, with GPU implementations showing an astounding **1000x** speedup in early tests. This allows the model to perform complex, real-time reasoning at scale.
## FAQ
**What is Ontology 1?**
Ontology 1 is a neurosymbolic search and discovery model developed by Onton. It is designed for complex, conversational, and multimodal product search by building an explicit, inspectable understanding of a user’s intent rather than relying on simple keyword or vector matching.
**How does it differ from Google or Amazon search?**
Unlike traditional search engines that primarily rely on keyword matching or statistical vector similarity, Ontology 1 actively reasons about a query. It identifies what it does not know, fills in logical gaps using its world model, and traces its decision-making process, leading to better results on complex, nuanced queries.
**Is there a public API or open-source version?**
No. Ontology 1 is not available as a downloadable model, API, or open-source checkpoint. Access is currently provided through the Onton.com platform, and partner access for building specific applications is granted on a case-by-case basis.
**What problems does it solve best?**
Ontology 1 excels at handling long, requirements-heavy queries where intent is ambiguous. It is particularly strong in verticals like home decor and furniture, where understanding abstract concepts like style, mood, and physical constraints is more important than matching specific product features.
**Is it multimodal?**
Yes, the model is designed for multimodal search. The benchmark tested text-only queries, but Onton reports that the model can handle image and multimodal inputs, although this specific evaluation was conducted on text-based searches.
## Conclusion
Ontology 1 represents a significant shift from data-driven search to reasoning-driven search. By building a symbolic, inspectable world model, it tackles the “black box” problem of modern AI and provides a compelling solution for complex e-commerce discovery. While its current index is small, the performance gains over industry giants and its novel architecture demonstrate a promising future for conversational commerce, where search engines act less like directories and more like knowledgeable shopping assistants.
