Perplexity.ai reinvents Google SEO model for LLM era

The release of ChatGPT on November 30, 2022 was met with much fanfare and a lot of backlash. It quickly became clear that people wanted to ask AI the same questions they asked Google, but ChatGPT often couldn’t answer them.

Problems were piling up. ChatGPT replies were outdated, did not cite sources, and frequently hallucinated new and inaccurate details. Emily Bender, director of the University of Washington’s Institute for Computational Linguistics, was quoted as saying at the time that AI search is “a Star Trek fantasy where you have an omniscient computer that you can ask questions of.”

Perplexity originally wanted to build an AI-powered Text-to-SQL tool. But something else started happening in his Slack channel at the company.

Founded in August 2022, Perplexity set out to build an AI-powered search engine that updates daily and cites multiple sources to answer queries. Ta. It currently has over 10 million monthly users and recently received an investment from Jeff Bezos.

“I think Google is one of the most complex systems humans have ever built. In terms of complexity, it probably exceeds flying to the moon,” said Perplexity.ai co-founder and CTO. says Denis Yarats.

First it was a Slack bot

Perplexity initially wanted to build an AI-powered Text-to-SQL tool to help developers write SQL queries and code in natural language, Yarats said. But something else started brewing in his Slack channel at the company. It’s a chatbot that combines search with OpenAI’s Large-Scale Language Model (LLM).

ChatGPT was then launched in late November 2022, becoming the fastest growing consumer application in history, reaching 100 million users within two months. People asked ChatGPT all sorts of questions, many of which could not be answered. But Yaratz says Perplexity’s Slack bot can do just that.

“In literally two days, we created a simple website, connected it to our Slack bot backend infrastructure, and released it as a fun demo,” says Yarats. “To be honest, it didn’t do very well. But given how many people liked it, I realized there was something there.”

For some time, Perplexity continued to develop Text-to-SQL tools. It also created BirdSQL, a Twitter search tool that allows users to search for very specific tweets, such as “Elon Musk’s tweets to Jeff Bezos.” But the AI-powered search engine stood out, and within months it became the company’s new and challenging mission.

How is AI-powered search possible?

This raises an obvious question. How did Perplexity, a company he founded less than two years ago with four people (and has since grown to about 40), solve the problems that seem to make AI terrible for search? Is not it?

Two decades of failures by Google’s competitors have proven that “decent” is not enough. AI provides a shortcut there.

Search Augmentation Generation (RAG) is one pillar of the company’s efforts. Invented by researchers at Meta, University of London, and New York University, RAG combines generative AI with a “retriever” that can search and reference specific data from a vector database, which is then passed to a “generator” that responds. will generate. .

“I also agree with RAG [is useful for search]” says Bob van Luijt, co-founder and CEO of AI infrastructure company Weaviate. “what [RAG] This makes it possible for regular developers, not just people working at Google, to build these types of AI-native applications without much effort. ” Resources to implement RAG are available for free on his HuggingFace for AI developer resources, he notes.

That led to widespread adoption. Weaviate uses her RAG to help clients establish AI agent knowledge based on their own data. Nvidia uses RAG to reduce errors in ChipNeMo, an AI model built to assist chip designers. Latimer uses it to combat racial bias and amplify minority voices. Perplexity then directs RAG to search.

However, to use RAG, your model must have the following conditions: something Here Perplexity.ai takes a more traditional search approach. The company indexes the Internet using his custom-designed web crawler known as PerplexityBot.

“If you want to perform well on breaking news and other news, you can’t retrain a model every day or even every hour,” Yarats says. However, crawling the web at Google’s scale is also not realistic. Perplexity lacks the resources and infrastructure of a tech giant. To manage load, Perplexity divides results into “domains” and updates them with more or less urgency. The news site he updates every hour he updates more than once. On the other hand, a site that is unlikely to change quickly will be updated once every few days.

Perplexity also leverages Bidirectional Encoder Representations from Transformers (BERT), an NLP model created by Google researchers in 2018. This model was used to better understand web pages. Google adopted BERT as open source, giving companies like Perplexity the opportunity to build on his BERT. “You get easy rankings. It’s not as good as Google, but it’s OK,” Yaratz says.

Keep Google at bay

But two decades of failures by Google’s competitors have proven that “decent” is not enough. AI provides a shortcut there.

“Google has a lot of constraints. The biggest one is advertising. The main page area is highly optimized.” —Denis Yarats, CTO, Perplexity.ai

LLMs excel at parsing text to find relevant information. In fact, finding patterns is kind of what the LLM is all about. This allows the LLM to generate persuasive text in response to prompts, but it can also be used to efficiently parse and present the information it examines. You can try this out for yourself by uploading the PDF to ChatGPT, Google Gemini, or Claude.ai. Within seconds, LLM can capture documents and answer questions about them.

Perplexity essentially does the same thing on the web, and in doing so fundamentally changes how search works. Rather than trying to rank web pages and place the best pages at the top of a list query, it analyzes the information available from an index of well-ranked pages to select the most relevant ones. Find and generate answers. That’s the secret sauce.

“You can think of the LLM as doing the final ranking task,” Yarats says. “[LLMs] please do not worry [SEO] Score. They only care about semantics and information. It’s fairer because it’s based on actual information rather than signals that Google engineers have somehow optimized. ”

Of course, this begs the question. Can’t Google do this too?

Yaratz said Perplexity recognizes the difficulty of taking on Google and is therefore focusing on “heads of distribution” for search. Perplexity does not provide image searches, cache old web pages, allow users to narrow results to a specific date or time, or include shopping results. Here are some other Google features that we often take for granted. He also believes Google will face problems related to its existing, profitable advertising business, rather than technical execution.

“Google has a lot of constraints,” he says. “The biggest thing is advertising. The main page area is highly optimized. You can’t say, let’s remove this ad, we’ll show you the answer instead. We don’t have that. We can experiment.”