Self-learning search

research

Nov 25, 2025

Knowledge Graph

Making sense of the increasingly unstructured web is critical to enabling high-quality search. And this sense-making can’t come from LLMs alone. As Meta’s Chief AI Scientist Yann LeCun said, “Auto-regressive LLMs [e.g. GPT] cannot be made factual, non-toxic, etc. ... It’s not fixable.”Sam Altman also discussed the limitations of large language models and that breakthroughs would be needed. What’s necessary is a solution that harnesses the power of models like GPT while taming their pitfalls.

To answer this, we’ve taken a novel approach. Similarly to how the two halves of the brain work together, we’ve engineered a powerful “left-brain” logic layer, backed by a fine-grained knowledge graph and reasoning engine, that extracts and makes explicit the “right-brain” statistical correlations encoded in LLMs. Our graph engine is 700x faster on natural language queries than RedisGraph (itself a state-of-the-art graph database which outperforms competitors by up to 15,000x on some benchmarks).

A knowledge graph is only as good as its data sources. We’ve developed a data pipeline capable of crawling, scraping, cleaning up, enriching, and indexing hundreds of thousands of products per day. After the initial ingestion, our tools automatically extract a wealth of untapped information that today’s big search engines overlook. This allows our users to find and make decisions more easily. We’re also planning to use this clean data to train and improve our own internal models.

Self-Learning

LLMs like GPT are trained to predict the next token. As they do this, they automatically construct an enormous world model sophisticated enough to support this prediction. In particular, the fluency of their linguistic production and the versatility of their problem-solving skills are nothing short of magical. But for all this sophistication, a large and growing body of research demonstrates the spottiness of their reasoning — not just on complex problems but even on seemingly basic ones.

There are at least three underlying problems that are fundamental to all LLMs. First, out-of-distribution generalization failure, which can manifest as “wonkiness”, “incoherence”, and “hallucination”. Second, a lack of iterativity, essential for multi-step reasoning, planning, and complex problem-solving. Third, training to token prediction rather than to truth. This last problem shows up everywhere LLMs are used, but is especially acute in e-commerce, where keyword monetization is too often more important than accuracy.

While working on a solution to the generalization problem to improve search accuracy, we ended up solving all three. The key to our solution is compositionality, by which small pieces of understanding are composed to form more complex ones. Our approach goes hand in hand with our name change to Onton — “onton,” meaning the smallest possible unit of being, beyond which distinction is impossible — and reflects our vision to “parse the web.”

So, how does our agent learn like a human? Think of two-year-olds learning language. They don’t do it by hearing millions of hours of speech from a gold-standard curated training corpus. They’re able to unconsciously construct increasingly accurate models of language from very little input — not by trial and error alone, but, for instance, by increasingly sophisticated “rules of thumb with exceptions”. And as linguistic ability continues to develop, children learn about the world by asking countless questions — the most precocious ones ask even more.

Our agent behaves similarly. Rather than making statistical sense of trillions of tokens via gradient descent, it learns compositionally. From a handful of axioms in set theory, we can derive not only all mathematical knowledge but everything mathematically knowable. Similarly, from simple beginnings, it learns increasingly sophisticated concepts. Our agent asks relevant questions about the world, performs thought experiments, and continually refines its knowledge.

All this means that the soon-to-be-released version of Onton learns from every single search. It doesn’t merely perform incremental training against an opaque, brittle statistical model. It performs targeted, automated, increasingly rigorous knowledge acquisition in direct response to user queries. Onton’s methodology yields, for the first time, truly trustworthy search.

Most recently, we’ve begun augmenting our knowledge graph via self-learning from queries and products. Soon, we’ll add user behavior and eventually multimodal data — images, videos, audio, and so on — to answer even more questions. Ultimately, we’ll have not just a hand-curated knowledge graph specific to product data, but a self-refining, self-correcting, self-augmenting one that encompasses all the background information and methods of reasoning needed to answer any query with precision.

From a business perspective, this will increase search accuracy and thus e-commerce conversion rate. Even small decreases in decision-making time or increases in conversion rate have a disproportionate impact on our growth, user retention, and revenue. We're anticipating much more than a slight increase, though, and we believe this is a 10x solution, much like what Page Rank enabled for Google to beat Yahoo. In addition, it means we could expand our market beyond e-commerce.

Research by: Alex Gunnarson

Knowledge Graph

Making sense of the increasingly unstructured web is critical to enabling high-quality search. And this sense-making can’t come from LLMs alone. As Meta’s Chief AI Scientist Yann LeCun said, “Auto-regressive LLMs [e.g. GPT] cannot be made factual, non-toxic, etc. ... It’s not fixable.”Sam Altman also discussed the limitations of large language models and that breakthroughs would be needed. What’s necessary is a solution that harnesses the power of models like GPT while taming their pitfalls.

To answer this, we’ve taken a novel approach. Similarly to how the two halves of the brain work together, we’ve engineered a powerful “left-brain” logic layer, backed by a fine-grained knowledge graph and reasoning engine, that extracts and makes explicit the “right-brain” statistical correlations encoded in LLMs. Our graph engine is 700x faster on natural language queries than RedisGraph (itself a state-of-the-art graph database which outperforms competitors by up to 15,000x on some benchmarks).

A knowledge graph is only as good as its data sources. We’ve developed a data pipeline capable of crawling, scraping, cleaning up, enriching, and indexing hundreds of thousands of products per day. After the initial ingestion, our tools automatically extract a wealth of untapped information that today’s big search engines overlook. This allows our users to find and make decisions more easily. We’re also planning to use this clean data to train and improve our own internal models.

Self-Learning

LLMs like GPT are trained to predict the next token. As they do this, they automatically construct an enormous world model sophisticated enough to support this prediction. In particular, the fluency of their linguistic production and the versatility of their problem-solving skills are nothing short of magical. But for all this sophistication, a large and growing body of research demonstrates the spottiness of their reasoning — not just on complex problems but even on seemingly basic ones.

There are at least three underlying problems that are fundamental to all LLMs. First, out-of-distribution generalization failure, which can manifest as “wonkiness”, “incoherence”, and “hallucination”. Second, a lack of iterativity, essential for multi-step reasoning, planning, and complex problem-solving. Third, training to token prediction rather than to truth. This last problem shows up everywhere LLMs are used, but is especially acute in e-commerce, where keyword monetization is too often more important than accuracy.

While working on a solution to the generalization problem to improve search accuracy, we ended up solving all three. The key to our solution is compositionality, by which small pieces of understanding are composed to form more complex ones. Our approach goes hand in hand with our name change to Onton — “onton,” meaning the smallest possible unit of being, beyond which distinction is impossible — and reflects our vision to “parse the web.”

So, how does our agent learn like a human? Think of two-year-olds learning language. They don’t do it by hearing millions of hours of speech from a gold-standard curated training corpus. They’re able to unconsciously construct increasingly accurate models of language from very little input — not by trial and error alone, but, for instance, by increasingly sophisticated “rules of thumb with exceptions”. And as linguistic ability continues to develop, children learn about the world by asking countless questions — the most precocious ones ask even more.

Our agent behaves similarly. Rather than making statistical sense of trillions of tokens via gradient descent, it learns compositionally. From a handful of axioms in set theory, we can derive not only all mathematical knowledge but everything mathematically knowable. Similarly, from simple beginnings, it learns increasingly sophisticated concepts. Our agent asks relevant questions about the world, performs thought experiments, and continually refines its knowledge.

All this means that the soon-to-be-released version of Onton learns from every single search. It doesn’t merely perform incremental training against an opaque, brittle statistical model. It performs targeted, automated, increasingly rigorous knowledge acquisition in direct response to user queries. Onton’s methodology yields, for the first time, truly trustworthy search.

Most recently, we’ve begun augmenting our knowledge graph via self-learning from queries and products. Soon, we’ll add user behavior and eventually multimodal data — images, videos, audio, and so on — to answer even more questions. Ultimately, we’ll have not just a hand-curated knowledge graph specific to product data, but a self-refining, self-correcting, self-augmenting one that encompasses all the background information and methods of reasoning needed to answer any query with precision.

From a business perspective, this will increase search accuracy and thus e-commerce conversion rate. Even small decreases in decision-making time or increases in conversion rate have a disproportionate impact on our growth, user retention, and revenue. We're anticipating much more than a slight increase, though, and we believe this is a 10x solution, much like what Page Rank enabled for Google to beat Yahoo. In addition, it means we could expand our market beyond e-commerce.

Research by: Alex Gunnarson

Knowledge Graph

Making sense of the increasingly unstructured web is critical to enabling high-quality search. And this sense-making can’t come from LLMs alone. As Meta’s Chief AI Scientist Yann LeCun said, “Auto-regressive LLMs [e.g. GPT] cannot be made factual, non-toxic, etc. ... It’s not fixable.”Sam Altman also discussed the limitations of large language models and that breakthroughs would be needed. What’s necessary is a solution that harnesses the power of models like GPT while taming their pitfalls.

To answer this, we’ve taken a novel approach. Similarly to how the two halves of the brain work together, we’ve engineered a powerful “left-brain” logic layer, backed by a fine-grained knowledge graph and reasoning engine, that extracts and makes explicit the “right-brain” statistical correlations encoded in LLMs. Our graph engine is 700x faster on natural language queries than RedisGraph (itself a state-of-the-art graph database which outperforms competitors by up to 15,000x on some benchmarks).

A knowledge graph is only as good as its data sources. We’ve developed a data pipeline capable of crawling, scraping, cleaning up, enriching, and indexing hundreds of thousands of products per day. After the initial ingestion, our tools automatically extract a wealth of untapped information that today’s big search engines overlook. This allows our users to find and make decisions more easily. We’re also planning to use this clean data to train and improve our own internal models.

Self-Learning

LLMs like GPT are trained to predict the next token. As they do this, they automatically construct an enormous world model sophisticated enough to support this prediction. In particular, the fluency of their linguistic production and the versatility of their problem-solving skills are nothing short of magical. But for all this sophistication, a large and growing body of research demonstrates the spottiness of their reasoning — not just on complex problems but even on seemingly basic ones.

There are at least three underlying problems that are fundamental to all LLMs. First, out-of-distribution generalization failure, which can manifest as “wonkiness”, “incoherence”, and “hallucination”. Second, a lack of iterativity, essential for multi-step reasoning, planning, and complex problem-solving. Third, training to token prediction rather than to truth. This last problem shows up everywhere LLMs are used, but is especially acute in e-commerce, where keyword monetization is too often more important than accuracy.

While working on a solution to the generalization problem to improve search accuracy, we ended up solving all three. The key to our solution is compositionality, by which small pieces of understanding are composed to form more complex ones. Our approach goes hand in hand with our name change to Onton — “onton,” meaning the smallest possible unit of being, beyond which distinction is impossible — and reflects our vision to “parse the web.”

So, how does our agent learn like a human? Think of two-year-olds learning language. They don’t do it by hearing millions of hours of speech from a gold-standard curated training corpus. They’re able to unconsciously construct increasingly accurate models of language from very little input — not by trial and error alone, but, for instance, by increasingly sophisticated “rules of thumb with exceptions”. And as linguistic ability continues to develop, children learn about the world by asking countless questions — the most precocious ones ask even more.

Our agent behaves similarly. Rather than making statistical sense of trillions of tokens via gradient descent, it learns compositionally. From a handful of axioms in set theory, we can derive not only all mathematical knowledge but everything mathematically knowable. Similarly, from simple beginnings, it learns increasingly sophisticated concepts. Our agent asks relevant questions about the world, performs thought experiments, and continually refines its knowledge.

All this means that the soon-to-be-released version of Onton learns from every single search. It doesn’t merely perform incremental training against an opaque, brittle statistical model. It performs targeted, automated, increasingly rigorous knowledge acquisition in direct response to user queries. Onton’s methodology yields, for the first time, truly trustworthy search.

Most recently, we’ve begun augmenting our knowledge graph via self-learning from queries and products. Soon, we’ll add user behavior and eventually multimodal data — images, videos, audio, and so on — to answer even more questions. Ultimately, we’ll have not just a hand-curated knowledge graph specific to product data, but a self-refining, self-correcting, self-augmenting one that encompasses all the background information and methods of reasoning needed to answer any query with precision.

From a business perspective, this will increase search accuracy and thus e-commerce conversion rate. Even small decreases in decision-making time or increases in conversion rate have a disproportionate impact on our growth, user retention, and revenue. We're anticipating much more than a slight increase, though, and we believe this is a 10x solution, much like what Page Rank enabled for Google to beat Yahoo. In addition, it means we could expand our market beyond e-commerce.

Research by: Alex Gunnarson