Andi, backed by Y Combinator, harnesses AI to create a better search engine • TechCrunch

It is difficult to convince users to switch search engines. This is one of the reasons why public search service startups rarely succeed. Another is that it is expensive to index a large number of websites (Google has a valued tens of billions of indexed pages), but a company backed by Y Combinator, And medon’t be discouraged – move forward to create an AI assistant that provides answers instead of links when searching online.

Andi was founded by Angela Hoover, who enrolled in YC’s Startup school after dropping out of college and entering YC’s 2022 winter batch. After working overseas in construction and with Microsoft as a data center project administrator, Hoover met Andi co-founder Jed White at the Denver airport upon returning to the States. -United.

Hoover and White — who had a background in artificial intelligence and search, specifically ranking, querying, and content quality classification — discussed how bad web search has become for things like travels and what it would take to create a new type of search engine from scratch.

“Gen Z hates Google. For us, search is interrupted. We live on our phones in messaging apps with visual feeds like TikTok and Instagram,” Hoover told TechCrunch in an email interview. don’t guess – Google executives have admitted “I hear my friends constantly say that Google sucks. Search results are overwhelmed with ads, SEO spam, and clutter. Gen-Z is so desperate for an alternative that we use TikTok as our search engine. We hate invasive creepy ads and the way Google is Big Brother and watches over everything.

Hoover offers Andi’s AI-powered assistant as an alternative. The general-purpose system attempts to find and extract answers to questions, combining large language models similar to OpenAI’s GPT-3 with live web data.

Behind the scenes, Andi pulls information from ranked web results based on relevance to the question being asked as well as overall quality (though it’s unclear how Andi defines “quality”). Depending on the topic, the platform uses different AI systems tailored to specific verticals (e.g. factual knowledge, programming, or consumer health) and language models that generate responses by combining knowledge from multiple sources (e.g. Wolfram Alpha, Forbes, The New York Times, etc.).

It’s a step beyond Google featured snippets, which extract text from web pages to answer frequently asked questions, and are akin to so-called “cognitive search” engines such as Amazon Kendra and Microsoft SharePoint Syntex which rely on knowledge bases to cobble together answers. Startups like Hebbia, Kagi, and You.com are also leveraging AI to return specific content from the web in response to queries, as opposed to just lists of results.

So what sets Andi apart? Unlike some of its competitors, Hoover says it does not charge for its service or log personally identifiable information. Andi also does not record or store the searches or results that people read or click on, only using coarse location data to improve the relevance of search results.

“Even when we add the option for user accounts in the future, we will only collect and retain enough data to help our customers use the service effectively, whether they want to create an account or be called back between devices and sessions, and to improve the service we provide,” Hoover said. “Users tell us that Andi can save them 15 or 20 minutes of searching, and have asked us to let them use it with their own team and personal data… As we improve question answering technology and add support for connecting to private data sources, we believe this has enormous potential.

To filter out information that might be misleading — or blatantly false — Hoover says Andi uses techniques like blocklists and ranking metrics. Misinformation is an ever-evolving problem, of course — a Google itself has struggle with. But Hoover expressed confidence in the technical measures taken by Andi to soften the impact.

“Every other new search startup is doing another weaker copy of Google with the same page cluttered with blue links meant for a web browser, with more or less variation on ads and privacy practices,” he said. she declared. “The content you see in [Andi’s] results are retrieved from the live source whenever possible, rather than from an outdated index. The response to questions is improving rapidly, and in many areas it is already excellent. »

In a quick experiment, I fed a few controversial queries into Andi and found that the search engine handled them quite deftly, consistently pointing to factual sources. A search for “Who really won the 2020 election?” gave the answer “Joe Biden”, while the question “Are COVID-19 vaccines fake?” pulled a Forbes article debunking pandemic conspiracy theories.

Andi is still very much in alpha and intends to stay lean as it iterates based on early user feedback, Hoover says. The startup will have difficult decisions to make. As New Yorkers Remarks, search algorithms are susceptible to various biases, such as prioritizing only websites that use modern web technologies. They also open the door to bad actors. In 2020, Pinterest took advantage a quirk of Google’s image search algorithm to make more of its content show up in Google image searches.

As it grapples with these issues, Andi’s team continues to figure out its business model. While the basic service will remain free, Hoover says Andi will eventually offer paid Professional and Commercial plans with premium features and API access, allowing customers to use Andi’s search and question answering capabilities with paid content, personal data, and internal company and team data. .

Paid features are probably the right way to go, given that Google’s share of the global search market has remained stable at over 90% for most of the past decade. Bing trails with 3.4%, followed by Yahoo! (full disclosure: parent company of TechCrunch) at 1.34%, according at Statcounter.

To fund the development of these features and potential partnerships with alternative search engines, Andi recently raised $2.5 million, including backing from YC, Gaingels, GoodWater Capital, K20 Fund, Acacia Venture Capital Partners, Fepo Capital and BBQ Capital, as well as a small family and friends around.

“We kept our burn rate low, working like digital nomads out of Mexico to expand our track and staying frugal. Even after adding AI developers and increasing our model training costs, we have well over two years of track,” Hoover said. “We’re using the funds to enhance our proprietary generative AI for answering complex questions, and the vertical search and API research technology Andi uses to combine large language models with live data, specifically: developing and training AI models, adding even more AI developers to our team and hosting and inference costs as we begin to increase usage once we get closer to product market fit… At this early stage, we are focused on creating great search that our users love, before generating revenue.”

Andi doesn’t collect detailed metrics, but Hoover estimates the search engine currently has around 5,000 users. Andi plans to add a full-time AI developer before the end of the year, bringing his total staff to three, including Hoover and White.