Korva is a scroll of the internet’s best ideas.
Agents aren't lazy searchers
Lazy human searches
Humans can hold a lot of words in a conversation. Typing a search is different: people lean on autocomplete to avoid typing three words, and settle for the shortest query that might work. Models are less lazy.
Cleaner agent queries
An agent's queries have fewer typos, are better specified, and may be longer, which leaves the search engine less to guess about what the agent wants. That opens up a different class of problems to solve.
A new interface
It also changed the interface. Serving agents instead of humans meant rethinking what an agent tells the search engine, because the more the engine knows about the request, the better it can do.
How does autocomplete shape what people type?
Google Suggest started in 2004 as a Google Labs project led by engineer Kevin Gibbs. It offered predicted completions as a person typed, based on common searches. Selecting a suggestion saves keystrokes, so people often accept it instead of finishing their own thought. Over time the suggestions nudge many searches toward the same popular phrasings.
How long are typical web searches?
A 1999 study of AltaVista query logs by Craig Silverstein and colleagues found that the average query had about 2.35 terms. Later studies by other groups found averages that stayed in the range of two to four words. Most people type the minimum that might work. Short queries leave a search engine to guess at what the person means.
Why do longer queries help search engines?
Many words have several meanings, and a single word like jaguar could mean an animal, a car, or a software product. Each extra term narrows the set of documents that could match. A query that spells out the context removes most of the ambiguity before ranking even starts. Engines have worked for years on query rewriting to add context that people leave out.
More from Training Data
