Perplexity initially used Bing as its search engine and OpenAI models to quickly build its first product.
“we initially used Bing, uh, as a as a search engine, and then uh, also obviously open aI models kind of like, uh, hacked quickly something.”
Person
Denis Yarats
1 episodes · 18 quoted claims
Perplexity relies heavily on open-source models such as Llama 3.
“we heavily, uh, rely on open source, um, things like Llama three is is very useful for us.”
Perplexity says it has been building its own search index for over a year, and the index was decent in size but smaller than it wanted.
“we've been investing this for like over a year now. Uh, it's uh. We have a a decent sized index”
Perplexity's ranking model targets whether a page's content will help answer the question, not the probability of a click.
“we're more more optimized on the if this content, if the if this web page was like, whatever snippets inside it has, is it is it gonna be helpful to answer the question?”
Perplexity builds trust scores for domains and web pages to prioritize high-quality sources and combat SEO spam.
“we spend a lot of time to build this like trust score for like Dom domain or or like a um, a web page.”
Yarats referred to a recent leak of Google's ranking signals as an example of how SEO practitioners learn what to manipulate.
“there was like, a couple of weeks ago, there was a a a leak, right?”
Yarats says Perplexity Pro users spread their model choices fairly evenly across the options offered.
“the distribution between like which models people select, uh, is, uh, quite uniform.”
Yarats estimated that at least 50% of Perplexity's traffic was served by in-house models.
“at least like 50% already served in house, uh, and like, yeah, on different models”
Perplexity post-trains Llama 3 70B, which Yarats described as the most capable open-source model at the time.
“we take like Llama three, um, like seventy B, that's sort of like the most capable open source model at the moment.”
Yarats said Perplexity receives roughly 10 million user queries per day.
“I think every day we might, I don't know, get getting like ten, ten, 10 million different prompts.”
Perplexity employs an internal team of annotators it calls LLM teachers, alongside third-party data vendors.
“we have a internal, uh, team of annotators, uh, we call them like, uh, Lm teachers.”
Perplexity uses human preference judgments to build a reward model, which is then used to train answer models.
“we get reward model, and then from the reward model, you train a model.”
Perplexity offers a default search mode and a Pro search mode that trade speed against depth.
“we right now on our product, we have like this two modes. One is just default search, the other one is like, pro search”
Yarats said Perplexity was started in July, before ChatGPT's release.
“we started the company, you know, in in July.”
Perplexity's first prototype, around October, was a Slack bot and Discord bot.
“around, like, um, I guess October, we had the the first prototype of Perplexity, where it was just like a Slack bot and Discord bot”
Yarats said swapping in OpenAI's Davinci three model without other code changes produced a large improvement, convincing the team the product could work.
“we like, literally replaced the model, um, without like any other code changes, and it was like, significantly.”
Yarats argued that users who click through from Perplexity have higher intent than those who click through from Google.
“when they go, they have like much higher, uh, intent, right?”
Yarats said the hardest part of scaling Perplexity was hiring good people.
“honestly, the hardest part was like hiring good people.”