Korva is a scroll of the internet’s best ideas.
Human click data is a bug
Cheaper feedback
Competing with the giants in search was hard partly because rivals lacked great feedback on which result beats another. Collecting it meant human ratings and click data at scale. Models change that, since ratings can now be created by experts far more cheaply.
Agents as customers
Models are also very good at compressing information, so a lot of the research that has gone into building them carries over to search indexing and ranking. The problem became more tractable because agents exist, both as technology and as a distinct customer with its own needs.
Growing the index
A company can then take on a billion-to-billion matching problem, with hundreds of billions of pages on one side and hundreds of billions of queries over time on the other, and grow its index step by step as it solves problems for more customers.
How do search engines use click data?
Search engines treat clicks as votes about which results satisfy a query. In a 2002 paper, Thorsten Joachims showed that clickthrough data could train a ranking system to order results. Clicks carry bias, though, because people click higher results more often just because they are higher. Engines build corrections for this position effect into their ranking models.
Why are model-made ratings cheaper than human ones?
A language model can judge thousands of results in the time it takes a human to judge a handful. A 2023 study by researchers at UC Berkeley and others found that GPT-4 agreed with human preferences more than 80 percent of the time on a chat benchmark. That is about as often as two humans agree with each other. Experts can then check a small sample instead of rating everything.
Why can't engines compare every query to every page?
Comparing every pair would take a number of operations equal to queries times pages, which is far too many to run in real time. Search systems turn text into embeddings and use approximate nearest neighbor methods to find close matches quickly. One widely used method, HNSW, was published by Yury Malkov and Dmitry Yashunin in 2016. It builds layered graphs so that a search can skip most of the data.
More from Training Data
