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

Neurosymbolic systems combine the exact logical reasoning of symbolic systems with the ability to learn from noisy data like neural networks. Traditional LLMs are good at approximation and robustness but are not optimized for factual accuracy and cannot perform genuine logical reasoning. The neurosymbolic approach applies exact answers where possible and learns from uncertain data where it cannot.

Vector search engines compare similarity vectors but cannot effectively understand subjective or complex criteria like style, room fit, or properties like "pet-friendly". They also lose information when transforming data into vectors and cannot verify whether product metadata is truthful.

Traditionally, humans manually create knowledge graphs, which is slow and limited to clear categories. Onton instead automatically expands the graph with relationships from real data and supports subjective predicates important for practical search, such as "cozy" or "won't look dated in ten years".

Ograph is a graph database optimized for fast querying of knowledge graphs. In 2023 it ran 700× faster than RedisGraph, and now outperforms SuiteSparse:GraphBLAS running on 14 cores with a single Ograph core, representing ~100× better throughput per core.