Perplexity AI

Perplexity AI is a retrieval-augmented answer engine founded in August 2022 by Aravind Srinivas (CEO, previously a researcher at OpenAI, Google, and DeepMind), Denis Yarats, Johnny Ho, and Andy Konwinski. Rather than returning a list of links, Perplexity synthesizes a written answer using a large language model and inlines numbered citations to the web pages it consulted. Under the hood the system runs a multi-stage pipeline: an intent parser breaks the user's question into sub-queries, a hybrid retriever combining lexical scoring (BM25) and dense embedding search fetches candidate pages, a multi-tier reranker scores them on factual density, structural clarity, and source authority, and a prompt assembler hands the model a packet with citations pre-embedded so the synthesized answer is constrained to the retrieved evidence. Typically around ten pages are retrieved per query and three to four end up cited in the visible answer. Perplexity does not train its own frontier model; it routes requests across third-party LLMs (variants of GPT, Claude, and open-weight models) and its own tuned variants. The product has expanded into a 'Pro' tier with deeper research modes, a built-in browser, and shopping features. As of September 2025 the company was reported as valued at about $20 billion, putting it among the most highly valued of the search-disruption startups, though it has also faced criticism from publishers over content reuse and from researchers over selective citation choices.