-
780m+
Queries/month
-
20%+
Monthly query growth
-
170M+
Monthly website visits
About Perplexity
Perplexity: Search-First AI
Perplexity is an AI-powered answer engine that combines web search with large language models to answer questions using current, cited information. Every response depends on retrieving the right information before generation begins.
Unlike traditional search engines, the goal isn't simply finding documents. The goal is finding the most relevant passages, ranking them effectively, and providing the language model with high-quality context to generate an accurate answer.
That retrieval layer is powered by Vespa.
First solve search, then solve everything elseAravind Srinivas, Perplexity CEO
The Challenge
Building an AI Answer Engine Starts with Better Search
Perplexity's goal wasn't simply to search the web. It was to generate accurate, grounded answers supported by citations. What separates a trustworthy answer from a plausible-sounding one is the quality of the information retrieved before the model ever runs. Perplexity built its product on that principle, which meant owning retrieval instead of renting it.
Building search in-house set a hard bar. The retrieval layer had to index a large, high-quality slice of the web, stay current as that web changed, and return results fast enough to feel instant, all while serving traffic at consumer scale.
When Perplexity evaluated existing tools, neither category cleared that bar. Perplexity found that vector databases could not reliably support even simple boolean filters, while traditional search backends offered only primitive semantic ranking.
The AI ecosystem had split into lexical systems and semantic systems, each strong on one side and weak on the other. Perplexity needed both, in one engine, at scale.
The team needed a retrieval platform capable of combining keyword and semantic search, filtering results, ranking individual passages, and continuously indexing new information—all within the latency expectations of a consumer search experience.
Rather than stitching together separate databases, vector stores, ranking services, and model-serving infrastructure, Perplexity looked for a single platform capable of performing retrieval and ranking as one integrated system.
The Solution
Why Perplexity Chose Vespa
Building an AI answer engine requires more than similarity search. To deliver accurate, grounded answers at the scale and speed users expect, Perplexity needed an AI search platform capable of finding, ranking, and serving the best information before generation begins.
Perplexity chose Vespa because it unifies the capabilities that AI search usually spreads across disconnected systems. Vespa unifies retrieval, ranking, filtering, real-time indexing, and machine learning inference in a single distributed serving engine. That gives Perplexity one platform for hybrid retrieval, passage-level ranking, real-time freshness, and integrated machine learning, without stitching together separate systems for each stage of the retrieval pipeline.
Bringing them together in a single distributed engine is what makes quality achievable at scale, and affordable to run.
What Great AI Search Depends On
Perplexity's approach to AI search is built around three core principles. Together, they define the architecture needed to retrieve accurate, trustworthy information at production scale.
Vespa: One Engine for Retrieval, Ranking and Inference
- Search at web scale. Continuously index billions of documents while serving thousands of low-latency queries per second, with content and computation distributed automatically across the cluster.
- Hybrid retrieval with structured filtering. Combine lexical, semantic, and structured signals within a unified retrieval and ranking pipeline, so results are both semantically relevant and exactly matched on names, terms, and filters.
- Passage-level retrieval and ranking. Rank and return the most relevant passages within documents, not just entire pages, giving the language model more focused, grounded context.
- Multi-stage ranking. Narrow large candidate sets efficiently with lightweight ranking, then apply richer neural models to the most promising results.
- Real-time updates. Continuously update content, vectors, metadata, and behavioral signals continuously, without interrupting live queries or rebuilding the index.
- Machine Learning in the serving layer. Execute ranking models and cross-encoders directly within the serving layer, eliminating external orchestration and its added latency.
- Cost efficiency at scale. Co-locating data, indexes, and computation on the same nodes avoids bandwidth bottlenecks and keeps infrastructure spend in check as traffic grows.
The Outcome
Perplexity Leads on Answer Quality
The same Vespa-powered retrieval platform that serves Perplexity's answer engine also powers its Search API. In benchmarks published by Perplexity, the platform demonstrated industry-leading speed, answer quality, and production scale.
- 358ms Median response time, making Perplexity's Search API the fastest commercial search API evaluated.
- 93% Top score across all four benchmark categories, including factual questions, multi-source reasoning, web research, and expert-level evaluations.
- 200M+ Queries served per day across an index of more than 200 billion URLs, continuously updated in real time.
One retrieval layer, two front doors
Perplexity's consumer answer engine and its Search API are not separate systems. Both query the same Vespa-based retrieval layer, so the benchmarks below reflect the same engine that answers questions on Perplexity itself. Vespa provides the retrieval, ranking, inference, and real-time indexing. Perplexity adds its web crawler, content understanding, and ranking models on top. The speed and scale figures come straight from the Vespa engine; the answer-quality scores reflect that engine working with Perplexity's models.
Building its own retrieval platform on Vespa enabled Perplexity to combine hybrid retrieval, machine-learned ranking, real-time indexing, and low-latency serving at web scale, which would have been difficult and costly to achieve by stitching together multiple systems.
How Vespa Delivers High-Quality Retrieval
Retrieve broadly, rank in stages, return only what the model needs.
Answering one question runs three stages in sequence, all on a live index that Vespa keeps current in real time.
1. Hybrid Retrieval
A single query pass combines vector similarity, lexical matching, and structured filters to pull a broad, precise candidate set from billions of documents. Vector search finds meaning; lexical search catches exact and rare terms like names and product codes; filters enforce recency and source.
More resources
Explore related guides, videos, case studies, and technical documentation in our resource library to learn more about AI search and retrieval with Vespa.
