Vibeleaderboard
Perplexity

ai lab

Perplexity

Perplexity matters because it made retrieval, citations, and answer generation feel like one consumer product, then turned the same search layer into developer APIs and a multi-model agent system. Its progress tests whether an AI-native search company can improve how people reach sources while building sustainable relationships with the publishers and communities whose material makes those answers useful.1,5,9,10

Official site ↗Models in the Atlas →

Profile

Overview

An answer engine becomes a company

Perplexity is a San Francisco answer-engine and agent company founded in 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski. Its original product replaced a conventional list of search results with a generated answer, inline citations, and suggested follow-up questions. The founders brought experience from OpenAI, Meta, Quora, and distributed-systems research, but the company's early distinction was a product architecture rather than a new general-purpose foundation model.1,2

Search infrastructure and Sonar

The company developed a search stack around retrieval, ranking, answer generation, and source citation. Sonar turned that stack into search-optimized models and APIs, while the Search API exposed ranked web results without requiring a generated answer. In 2025 Perplexity also published search_evals, a reproducible framework for comparing search APIs on single-step and deep-research tasks. These releases made parts of the system inspectable even though the production index and full ranking pipeline remain proprietary.4,5,6,7

From search to agent orchestration

Perplexity then widened from question answering into the Comet browser, Deep Research, enterprise search, and Computer. Computer assigns subtasks across several external and internal models, tools, files, and web services instead of presenting one model as the whole product. Portable Computer extended that orchestration to supported local NVIDIA hardware, with explicit permission required before a step escalates to a cloud model. The strategy makes Perplexity a model router, search provider, and agent product at the same time.8,9,12,13

The publisher conflict

That reliance on the open web also created the company's central dispute. Forbes, Reddit, and other publishers or platform operators accused Perplexity or its data suppliers of reproducing reporting or bypassing technical controls, while Perplexity argued for broad access to public knowledge and also signed licensing arrangements with some publishers. In July 2026 a federal judge allowed most of Reddit's scraping case to proceed. The litigation has not resolved whether the underlying conduct was unlawful, but it makes source governance and publisher economics material to any assessment of the company.2,10,11

Notable contributions

  1. 01Popularizing cited conversational searchPerplexity placed generated answers, inline source links, and follow-up questions at the center of the search experience. It did not invent retrieval-augmented generation, but it helped turn cited answer engines into a recognizable consumer category.1,2,3
  2. 02An inspectable search evaluation frameworkThe company exposed ranked web retrieval separately from answer generation and released search_evals with provider harnesses, cost accounting, traces, and benchmark integrations. The contribution is a reproducible evaluation tool, not proof that Perplexity's own reported scores are neutral.5,6,7
  3. 03Orchestration across specialized models and toolsComputer routes work across specialized models and tools, while Portable Computer starts supported tasks on local hardware and asks before cloud escalation. These releases make deployment location and model choice explicit parts of Perplexity's agent design without claiming that the company invented multi-model orchestration.8,9,12,13