Perplexity

What Perplexity is and how it organises search

Perplexity is an AI and search platform that retrieves information and writes referenced answers. Its offering also includes research, file creation, the Comet browser and task-performing agents. Distinguish these experiences before choosing a workflow.

Conversation helps refine questions and follow references. Value includes finding documents that can be checked. Source selection and answer generation are separate steps, both capable of error.

For OSH, useful output is a clearly scoped documentation trail back to the organisation, document and passage supporting a claim, without treating the synthesis as an independent authority.

Reference documentation: Perplexity: funcionamiento de Pro Search · Perplexity: presentación de Computer.

Origins, founders and company development

Perplexity began in 2022, founded by Aravind Srinivas, Denis Yarats, Johnny Ho and Andy Konwinski. NEA’s Srinivas interview describes conversational search as a route to better knowledge access.

In January 2024 it announced a $73.6 million Series B led by IVP. This is a historical funding event, not a current valuation or reliability measure.

Deep Research and Labs expanded the offering in 2025; Computer added coordinated task work in 2026. Interpret older product names within that chronology.

Reference documentation: NEA: inversión y equipo fundador de Perplexity · NEA: entrevista con Aravind Srinivas sobre el origen de Perplexity · Perplexity: ronda serie B de enero de 2024 · Perplexity: lanzamiento de Labs.

Search, Pro Search and model selection

Pro Search combines searches, synthesis and references, with model and focus options depending on plan. It uses its own and third-party models; this is not a subscription to those providers’ original applications.

Match the mode to the question. Finding one document needs precision; comparing alternatives needs separate criteria, evidence and conclusions. Length does not fix poor scope.

Retain prompt, date, mode and visible model. Compare the same questions. Earlier tests remain evidence of that configuration, not a permanent guarantee after model changes.

Reference documentation: Perplexity: funcionamiento de Pro Search.

Research: deeper investigation and traceability

Deep Research launched in February 2025 with iterative search, reading and reporting. Later documentation also uses Research. Check current limits and models.

Define activity, country, period and evidence type. Separate official documents, studies and commentary. Open originals before adopting conclusions.

Preserve differences in population, method and date. Include them and missing data in a comparison matrix. Narrative coherence should not erase disagreement.

Reference documentation: Perplexity: lanzamiento de Deep Research · Perplexity: lanzamiento de Labs.

Model Council and comparing responses

Model Council consults three models and synthesises agreement, differences and contributions. It has also reached Computer, subject to product and subscription access.

Comparison can expose differing interpretations. Agreement does not establish correctness; models may share sources or errors.

Use differences to identify unresolved questions and check primary documentation. Record acceptance reasons beyond a model vote.

Reference documentation: Perplexity: Model Council y evolución de las modalidades · Perplexity: Projects y Personal Computer para Windows.

From Labs to creating files and apps

Labs launched in May 2025 to produce reports, spreadsheets, charts and simple apps through research and code. The interface adopted Create files and apps.

Check the actual file: formulas, units, omissions and totals; presentation sources and readability; app inputs, permissions and persistence.

Use public or fictional data in OSH prototypes. Define ownership, maintenance and error handling before operational use. A generated interface does not establish readiness.

Reference documentation: Perplexity: lanzamiento de Labs · Perplexity: Model Council y evolución de las modalidades.

Computer: multi-step tasks

Computer was announced on 25 February 2026 to coordinate models and tools. It can produce deliverables and act on files and services within permissions and configuration.

Define outcome and verification: sources, period, structure and audience. Distinguish drafting from external modification or publication.

Accept through opened files, reconciled data, valid links and visible changes. A completion message does not replace verification.

Reference documentation: Perplexity: presentación de Computer.

Projects, shared files and memory

Spaces evolved into Projects with persistent files, conversations, connectors and memory. Collaboration does not imply inheriting another person’s private connections.

Group approved sources and drafts around a topic, with current versions, owners and retirement rules. Memory should not make earlier interpretations immutable.

Test visibility and editing before invitations. Outputs derived from restricted sources may need equivalent access limits.

Reference documentation: Perplexity: Projects, evolución de Spaces.

Skills, connectors and recurring work

Skills reuse instructions, connectors provide service access, and automations repeat runs. Review each component separately.

An OSH example is a periodic draft of updates from defined organisations. Specify sources, period, inclusion and uncertainties; publication requires editorial judgement.

Test missing sources, changed structure and no-news periods. Do not invent material to fill templates. Distinguish no change from retrieval failure.

Reference documentation: Perplexity: funciones y recorridos de Computer · Perplexity: esfuerzo, modelos y ejecución local, septiembre de 2026.

Teammates: agents for ongoing work

Perplexity introduced Teammates on 4 November 2026 for ongoing tasks, sessions and agent coordination within Computer. Initial launch was web for Max, with iOS planned later.

Define initiation authority, review requirements and stop conditions. Document monitoring can return changes and proposals without modifying current procedures.

Review open tasks and accumulated assumptions. Corrections should carry into subsequent work and remain findable.

Reference documentation: Perplexity: Teammates, 4 de noviembre de 2026.

Local, hybrid and cloud execution

Computer distinguishes cloud, hybrid and local modes. October 2026 brought fully local execution on supported Macs: model, planning and files can remain on-device, with permission for cloud escalation.

Hybrid mode divides work between cloud and local models with outgoing-data controls. It is not wholly offline use. App location does not establish processing location.

Identify components and destinations. Test network interruption or closure with fictional data; record persistence, resumption and outgoing information.

Reference documentation: Perplexity: Computer local en Mac, octubre de 2026 · Perplexity: procesamiento híbrido en Mac.

Comet: browser and assistant

Comet is Perplexity’s browser, introduced in July 2025. Its assistant uses browsing context and can act. Distinguish it from Search and Computer.

Web content is third-party information, not authorised instruction. BrowseSafe++ studies malicious instructions inserted into that context.

Limit scope and verify changes, sends or downloads. Start with public-page comparison and drafts; review browser permissions as well as answers.

Reference documentation: Perplexity: presentación de Comet · Perplexity: investigación BrowseSafe++.

APIs, integrated search and development

Documentation distinguishes Agent API, Search API, Orchestrator, Embeddings and Decisions. Search returns results and text; Agent can generate cited answers. Choose retrieval or generation according to need.

Retain query, date, references and errors. Mark unretrieved documents rather than implying they were read.

Test limits, cost, retries and versions. Keep credentials out of visible logs. API availability does not establish a ready-made integration.

Reference documentation: Perplexity: catálogo de APIs.

Privacy, memory and enterprise data

Individual accounts can opt out of model-improvement contribution through settings. Enterprise plans publish no-training conditions for organisational data. Review the mode and contract.

Memory, history and training differ. Brain has its own controls and is documented as Research Preview. Disabling one function does not necessarily delete another’s data.

Classify and minimise workplace data. Review inputs and outputs; context can identify people even after names are removed.

Reference documentation: Perplexity: memoria y personalización · Perplexity: controles de Enterprise Max · Perplexity: Max y Brain en Research Preview.

Plans, consumption and professional selection

Perplexity distinguishes Free, Pro, Max, Education Pro, Enterprise and API access, with feature-specific limits. The former Organization File Repository retired on 1 October 2026; it is not a new available entitlement.

Define volume, collaboration, controls and outputs. Individual payment does not imply enterprise administration. API usage is separate.

Compare cost and review effort using representative tasks. Measure time to a verified answer, including retries for omissions.

Reference documentation: Perplexity: planes, modalidades y privacidad.

Publishers, attribution and controversies

Perplexity created a publisher programme in 2024. Content use also drew controversy: News Corp reported Dow Jones and New York Post proceedings; Perplexity rejected their characterisation in October 2024.

In August 2025 Cloudflare published tests it interpreted as automated access contrary to restrictions. Perplexity disputed attribution and distinguished crawling from user-triggered requests. These are documented party positions, not a final judicial determination asserted here.

Preserve attribution and review reuse rights. Finding or summarising a source does not grant full reproduction rights. Provide original synthesis and links to originals.

Reference documentation: Perplexity: programa de editores · News Corp: comunicación de procedimientos contra Perplexity, noviembre de 2024 · Perplexity: respuesta sobre la demanda de Dow Jones · Cloudflare: investigación de tráfico de Perplexity, agosto de 2025 · Perplexity: respuesta a Cloudflare.

Applications in occupational safety and health

It can help locate guidance, reports and reference materials about occupational safety and health. Its preventive value lies in finding useful documents more quickly, followed by checks on authority, currency, jurisdiction and relevance to the activity concerned. Search results remain starting points for professional review.

Beyond finding guidance, it can map organisations publishing on a topic, identify documents updating earlier recommendations, or compare how different sources define an indicator. These are documentary workflows proposed by Sabentis. A particularly useful output separates primary sources, commentary and commercial material. If a response mixes all three, reconstruct the source table before transferring its statements into a procedure. Also identify whether apparently different documents rely on the same underlying study or announcement.

Worked example and checks

For example, a practitioner can search for official guidance on heat exposure at a workplace. They open the cited documents, check dates and scope, select applicable material and compare possible measures before deciding what to include in the prevention plan.

To develop the heat-exposure search:

1. Define the activity, country and need: finding relevant official documents, without initially requesting a universal recommendation.
2. Request title, organisation, date, link and the population or activity covered. Ask for unclear dates or scope to be flagged.
3. Open every selected document. Check whether it is guidance, a legal instrument, an information note or a superseded version.
4. Build a matrix of comparable topics: work organisation, information, incident response and assessment requirements. Do not equate categories that the sources define differently.
5. Record exclusions and unresolved questions. A competent person then decides how the evidence relates to the site.

The deliverable should let a reviewer locate the passage supporting each row and preserve the access date. Where two sources disagree, present the difference and its scope; do not resolve it through an average or a synthesis that hides the disagreement. Retain the original wording of key definitions in working notes, with appropriate attribution, so changes in meaning can be checked.

Limitations and conditions of use

A citation does not guarantee that the source supports the summarized statement or remains current. Answers may mix jurisdictions and evidence quality; decisions should rely on reading and evaluating the originals.

Automatic selection may omit relevant material or favour readily accessible documents. For an important search, check known organisations and repositories even if they do not appear in the initial answer. A question containing an implied conclusion can steer results towards confirming it; use a second query looking for conditions, exceptions or contrary evidence. If a reference cannot be opened, mark it as awaiting verification rather than reviewed. A publication date is also different from the date of the evidence it discusses.

Choosing and complementing the tool

Perplexity fits web exploration that needs traceable links. When a question requires extracting variables from scientific studies, it can complement Elicit. For repeatedly querying an already selected document collection, compare the workflow with Gemini Notebook.

A pilot should assess relevant documents found, rejected sources, coverage of key organisations and citation accuracy. Counting links alone does not measure quality: several pages may repeat the same statement without adding independent evidence. Keep the search scope stable when comparing approaches so differences are meaningful.

Frequently asked questions

Do citations turn an answer into a systematic review?

No. A systematic review needs a reproducible search, selection and appraisal protocol. An answer with links may support part of the work but does not demonstrate that process.

Is an answer with more sources better?

That depends on relevance, independence and quality. One pertinent primary source may be more useful than many pages repeating it.

How should research be retained?

Save the question, date, selected documents, exclusion reasons and references to specific passages. Another person can then review the reasoning without relying on a previous conversation.