Onton Releases Ontology 1: A Neurosymbolic Search Model That is 2.7x More Accurate than the World's Best E-commerce Search Engines
Onton released Ontology 1, a neurosymbolic search model for complex, conversational, multimodal product search that outperforms Google Shopping and Amazon on a 90-query benchmark Ontology 1 achieved a mean precision@10 of 0.630 versus 0.543 for Google Shopping and 0.469 for Amazon, while indexing only ~1% of their catalogs The model uses an inspectable knowledge graph (Ograph) that decomposes vague predicates like "pet-friendly" into checkable properties (fiber, weave, durability) rather than re
Analysis
TL;DR
- Onton released Ontology 1, a neurosymbolic search model for complex, conversational, multimodal product search that outperforms Google Shopping and Amazon on a 90-query benchmark
- Ontology 1 achieved a mean precision@10 of 0.630 versus 0.543 for Google Shopping and 0.469 for Amazon, while indexing only ~1% of their catalogs
- The model uses an inspectable knowledge graph (Ograph) that decomposes vague predicates like "pet-friendly" into checkable properties (fiber, weave, durability) rather than relying on seller labels
- The benchmark (Subtext-Decor-90) was scored by three independent LLM judges (Claude Opus 4.8, Gemini 3.1 Pro, GPT-5.5), with modest inter-judge agreement (Krippendorff's alpha 0.465) but consistent ranking across all judges
- The model is not open-weight; it is available only as a live product on Onton.com with partner access granted case-by-case, targeting mid-market and enterprise retailers
Why It Matters
This represents a significant step toward neurosymbolic AI in production search systems, demonstrating that explicit reasoning over knowledge graphs can outperform pure vector/keyword retrieval on complex, requirements-heavy queries. For AI practitioners, it validates the growing interest in hybrid architectures that combine neural and symbolic approaches, particularly for domains where interpretability and trustworthiness matter. The results also highlight the limitations of current e-commerce search interfaces, which have remained largely unchanged for decades despite advances in AI.
Technical Details
- Architecture: Neurosymbolic model combining an inspectable knowledge graph (Ograph, a custom graph database) with neural components. The system builds explicit world models rather than absorbing patterns solely into weights, enabling reasoning from objective properties (e.g., fiber, weave, construction) to infer answers to vague queries.
- Benchmark: Subtext-Decor-90, a 90-query text-based benchmark with code and data released publicly. Three multimodal LLM judges (Claude Opus 4.8, Gemini 3.1 Pro, GPT-5.5) scored precision@10 across top result cards. Results: Onton 0.630 [0.571, 0.688], Google Shopping 0.543 [0.490, 0.596], Amazon 0.469 [0.417, 0.521], with 95% confidence intervals from 10,000 bootstrap resamples.
- Performance: Onton won 52 of 90 queries outright. When excluding empty result slots (3 queries), Onton's P@10 rises to 0.665. The model indexes only ~1% of competitor catalogs yet still outperforms them.
- Infrastructure: Ograph custom graph database demonstrates strong performance — one core beats SuiteSparse:GraphBLAS on 14 cores (~100× throughput per core), with GPU build running 43× faster than CPU and early runs approaching 1000× as implementation is tuned.
- Limitations: Failure cases occur on functional-spec queries where Amazon's category metadata dominates (e.g., "lamp that won't wake my partner" — Onton 0.4 vs Amazon 0.9). The model currently indexes only home decor/furniture, though Onton claims the methodology generalizes to non-product data with minimal reconfiguration.
- Availability: No public API, pricing tier, or open checkpoint. Access is through partnerships for agentic web teams.
Industry Insight
- The neurosymbolic approach validates a growing industry trend: pure neural methods struggle with complex, multi-constraint queries where interpretability and reasoning transparency are critical. Companies building search or recommendation systems should evaluate hybrid architectures, especially for domains with high-stakes or ambiguous user intent.
- The benchmark methodology — using multiple LLM judges with reported confidence intervals and inter-judge reliability metrics — sets a useful standard for evaluating search systems. Practitioners should adopt similar rigorous evaluation practices rather than relying on single-metric or single-judge assessments.
- The partnership-only distribution model suggests that competitive advantage in this space comes from proprietary data and graph infrastructure rather than model weights alone. Organizations should consider building domain-specific knowledge graphs as strategic assets, particularly for verticals with complex product attributes and noisy listing data.
Disclaimer: The above content is generated by AI and is for reference only.