Perplexity

Overview

Perplexity is an AI search company focused on combining retrieval, ranking, and model inference into a production search experience. Rather than treating search as a simple wrapper around a language model, Perplexity appears to operate as a tightly integrated system: embeddings help rank and retrieve relevant sources, inference servers generate or refine answers, and batching plus latency optimization keep the experience fast enough for consumer and enterprise use.

For AI Product Managers, Perplexity matters because it is a concrete example of what it takes to ship reliable, real-time AI search at scale. Its recent updates span search infrastructure, agent APIs, SDKs, custom skills, licensed data access, and research-oriented products such as Perplexity Computer. That makes it useful as a case study in production retrieval-augmented systems, multi-model orchestration, cost/latency trade-offs, and the packaging of internal AI capabilities into developer platforms.

Key Developments

• 2026-07-18: Perplexity’s Agent API added support for custom skills, along with documentation and a research PDF cookbook to help developers build and integrate new skills.
• 2026-07-26: Perplexity unveiled a CLI for live web data.
• 2026-07-29: Aravind Srinivas announced that Perplexity open-sourced Bumblebee, a client-side read-only scanner agent for detecting vulnerabilities on macOS and Linux.
• 2026-08-01: Perplexity CISO kpolley open-sourced Numbat, a real-time monitoring tool designed to detect and manage risky AI coding agents.
• 2026-08-14: Perplexity’s Wide-And-Deep-Research benchmark, using the Perplexity Computer harness, was used to benchmark Grok 4.6 as an orchestrator; it landed on the performance-versus-cost Pareto frontier and became available to Perplexity Pro and Max users.
• 2026-08-15: Perplexity released its Search SDK for use inside agentic harnesses, described as core technology behind Perplexity Computer’s wide and deep research capabilities.
• 2026-08-21: Aravind Srinivas positioned Perplexity’s Agent API as a full AI developer platform, offering access to frontier and workhorse models plus tooling for production deployment.
• 2026-08-27: Perplexity Computer added connections to Dun & Bradstreet, Guidepoint, IBISWorld, and 20+ other licensed data sources, enabling users to query firm-licensed information with traceable source records and without separate APIs or logins.
• 2026-08-29: Aravind Srinivas shared that Decagon uses Perplexity search to provide online information for customer-support-related questions in deployments serving enterprises such as Delta Airlines, Ticketmaster, Deutsche Telekom, and American Airlines.
• 2026-09-05: Aravind Srinivas published a deeper look at how Perplexity serves search results at scale, covering embeddings for ranking, GPU-based model inference, request batching, inference servers, and latency/throughput trade-offs.

Relevance to AI PMs

1. A practical blueprint for production AI search: Perplexity shows that good AI search is a systems problem, not just a model problem. PMs can study its use of embeddings, ranking, inference infrastructure, and batching to understand how to design products that are both accurate and fast.

2. Useful model for developer-platform packaging: With its Agent API, Search SDK, CLI, and custom skills support, Perplexity demonstrates how a company can turn internal AI capabilities into reusable platform primitives. PMs building AI platforms can use this as a template for deciding what to expose as APIs, SDKs, and workflows.

3. Strong example of research and enterprise workflow integration: Perplexity Computer’s connections to licensed sources and traceable source records highlight a practical path for AI products serving analysts, support teams, and knowledge workers. PMs can apply these ideas when prioritizing source attribution, enterprise access controls, and high-trust retrieval.

Related

• aravind-srinivas: CEO and a primary public voice explaining Perplexity’s product direction, infrastructure, and platform strategy.
• perplexity-computer / computer: Perplexity’s research-oriented product and harness layer for wide and deep research tasks.
• perplexity-agent-api / agent-api: The company’s developer platform for building production AI workloads with models and tools.
• search-sdk: SDK layer that exposes Perplexity search capabilities for agentic systems and harnesses.
• pplx-embed / embeddings: Related to Perplexity’s ranking and retrieval stack, especially in search relevance workflows.
• gpu-based-model-inference / nvidia-ai / cutlass-python-stack: Closely connected to Perplexity’s infrastructure story around inference efficiency, throughput, and latency.
• deep-and-wide-research / agentic-harnesses / computer-skill-memory: Related concepts and tooling around orchestrated research workflows.
• decagon: Example of Perplexity search being embedded into enterprise customer support experiences.
• claude, openai, anthropic, google, gemini, grok-45, grok-46, mistral-ai, kimi-k25, nemotron-3-ultra: Model and platform ecosystem peers or dependencies relevant to Perplexity’s multi-model positioning.
• samsung, apple, amazon, meta, microsoft, reddit, wikimedia-foundation: Broader platform, distribution, content, or ecosystem entities that may intersect with Perplexity’s search and agent strategy.
• robotstxt, geo, browsesafe: Related to web access, indexing, content governance, and safe browsing considerations for AI search products.