Korva is a scroll of the internet’s best ideas.

Why search exists

Search exists because the web grew from thousands of pieces of content to millions, billions, and trillions, far too many for a human to sort through. People needed a layer where they could ask for what they were looking for and get back reasonably relevant answers. Google came up in an era when the main agent acting on the web was a human, so it was tuned to what humans need, and tuned to sell a lot of ads.

What agents need

Agents have different needs, and different content is optimal to serve them. Even though models were trained on a big slice of the internet, the internet keeps changing, and search is how a model reaches fresh information it was never trained on.

Exa's search API

Exa was one of the first to build an AI-specific search engine and API optimized for what agents do. Its founders started it before ChatGPT took off, and it is now one of the go-to options for agents that search the web.

How does AI search differ from keyword search?

Classic keyword search ranks pages by matching the words in a query, with methods such as BM25. Embedding-based search turns text into numeric vectors and ranks pages by closeness of meaning. That lets an agent describe what it wants in a full sentence and still get matching pages. Many AI-oriented search engines combine both approaches.

Why can't a model just memorize the web?

A model's knowledge is frozen at its training cutoff, so anything published later is unknown to it. A 2020 paper from Facebook AI Research introduced retrieval-augmented generation, in which a model fetches relevant documents at answer time. Even for older material, models compress their training text and can recall details inaccurately. Search gives an agent the original page to read.

How do agents search differently than people?

A person types a few words and scans a page of results. An agent can write long, specific queries and issue many of them at once. It reads full page text, so it benefits from content with ads and navigation stripped out. Search APIs built for agents typically return page contents directly.

More from Lightspeed Venture Partners

Tech & AINothing got replaced, only layeredPrompt, context, harness, and loop engineering stack on top of each other, and every layer still changes how an agent behaves.
Enjoyed it?