Korva is a scroll of the internet’s best ideas.
A thousand tokens from a trillion pages
Routing the query
A query is first run through models that work out what it means and enrich it. It is then routed to several indexes that organize the web in different ways: a big index, a fresh index, and structured or knowledge-graph style ones. The query is rewritten for each.
Ranking layers
Each index has a retrieval layer followed by several ranking layers. Hundreds of billions of URLs shrink to thousands, then tens of thousands, then down to specific excerpts and paragraphs, with bigger models and different architectures at each stage pulling more features.
What comes out
What comes out is the thousand tokens with the highest signal for the AI. Less noise in the results gives the model room to do more: searches built this way use under half the tokens in an agent and come out more accurate and faster. For a model limited by context or memory, that means more problems solved, or the same ones solved cheaper.
Different versions of the search API spend different amounts of compute at various points in that journey. Until last week, the compute at read time had three seconds to work with. A new product does it in 200 milliseconds.
Why does one query go to several indexes?
Different parts of the web change at different speeds. News sites update by the minute, while reference pages may stay the same for years. Search engines have long split their data into tiers so that the fast-changing tier can be rebuilt often. Splitting the work also lets each index use the structure that suits its content.
Why do ranking systems use bigger models later?
Large models cost too much to run on billions of candidates, so engines run cheap models first to cut the list. Later stages apply more expensive models to the few results that remain. This design is called cascade ranking, and Lidan Wang, Jimmy Lin, and Donald Metzler described it in a 2011 paper. Each stage spends compute only where it can still change the outcome.
Why does less noise help a language model?
Researchers at Stanford and other institutions found in 2023 that models use information worse when it sits in the middle of a long input. That finding was published as Lost in the Middle. Extra text also costs money, because providers charge by the token. Fewer, better excerpts leave the model more room to reason.
Why does 200 milliseconds matter?
Usability researcher Jakob Nielsen wrote that a response under about 0.1 seconds feels instant to a person. A software agent has no such perception, but it chains many searches in sequence, so each delay adds up. Cutting a search from seconds to a fraction of a second shortens every step of the chain. Faster steps let the agent run more searches in the same time.
More from Training Data
