Perplexity

The documentation explains how to create API keys and organize them into API groups for different projects or environments, and shows that Search, Chat, Embeddings, and Agent endpoints all use Bearer token authentication, but it does not specify whether an API key can be restricted to only the Search endpoint or whether keys provide full access to chat, embeddings, and agent features by default.

Add explicit documentation on the API Key Management page clarifying whether API keys can be scoped to individual endpoints or if every key grants universal access to Search, Chat, Embeddings, and Agent APIs by default.

Where we looked: /docs/admin/api-key-management.md, /docs/admin/api-key-management, /api-reference/embeddings-post, /api-reference/agent-post, /api-reference/search-post, /api-reference/sonar-post

The documentation does not explicitly list any unsupported parameters or features. It only enumerates supported standard OpenAI parameters—for the Sonar API: `model`, `messages`, `max_tokens`, `stream`, `temperature`, `top_p`, and `response_format`; for the Agent API: `model`, `input`, `instructions`, `max_output_tokens`, `stream`, and `tools`—plus Perplexity-specific additions. No limitations or excluded OpenAI parameters are mentioned.

Add an explicit 'Unsupported Parameters' or 'Limitations' section to the OpenAI compatibility pages that lists which OpenAI parameters and features (e.g., logprobs, function calling, vision, etc.) are not supported.

Where we looked: /docs/sonar/openai-compatibility, /docs/sonar/openai-compatibility.md, /docs/agent-api/openai-compatibility

The documentation does not explicitly state whether rate limits apply per individual API key or across the entire organization, though it notes that usage tiers are based on “cumulative purchases across your account lifetime” [4] and that Search API limits “apply consistently across all accounts” [1]. It also does not specify whether limits reset on a rolling window or at a fixed calendar time; the Search API is described as using a “leaky bucket algorithm” [1], while Sonar and Agent APIs list QPS and RPM limits [2][4] without describing reset behavior.

Add explicit documentation on whether rate limits are enforced per API key or aggregated across the organization, and clarify whether QPS/RPM limits reset on a rolling window or at fixed calendar intervals.

Where we looked: /docs/admin/rate-limits-usage-tiers.md, /docs/admin/rate-limits-usage-tiers.md, /docs/admin/rate-limits-usage-tiers.md, /docs/admin/rate-limits-usage-tiers.md, /docs/resources/faq

Perplexity's documentation states a strict Zero Data Retention Policy for the Sonar API, meaning it does not retain any data sent via the API beyond processing the immediate request. However, the retrieved chunks do not specify a retention TTL or how long async chat completion results remain available before deletion.

Add documentation for the async chat completions endpoint that specifies how long results are retained, the polling window, and any hard TTL before deletion.

Where we looked: /docs/resources/privacy-security, /docs/getting-started/integrations/pipedream, /docs/getting-started/integrations/pipedream.md

The API reference for `generate_auth_token` shows the request body only accepts an optional `token_name` parameter and does not include an expiration field, indicating custom expiration is not supported. The documentation does not specify a default validity period for auth tokens. The troubleshooting section mentions checking "API tier limits" if key generation fails, but it does not state a maximum number of active tokens allowed per account.

Document the default auth token validity period, whether custom expiration can be configured, and the exact maximum number of active tokens per account on the API Key Management or Generate Auth Token reference pages.

Where we looked: /api-reference/generate-auth-token-post.md, /api-reference/generate-auth-token-post, /api-reference/generate-auth-token-post, /docs/admin/api-key-management, /docs/admin/api-key-management

AWS Marketplace subscribers appear to receive the same model access and pricing as direct API customers: purchased credits are applied to your Perplexity API balance and work across all APIs (Sonar, Agent API, Search API, and Embeddings), and credits are denominated 1:1 with USD, drawn down according to the standard per-request and per-token pricing. However, the retrieved chunks do not state whether rate limits and usage tiers are identical for AWS Marketplace customers compared with direct API customers.

Add an explicit statement on the AWS Marketplace page confirming whether rate limits and usage tiers are the same as direct API customers, or list any differences.

Where we looked: /docs/resources/aws-marketplace, /docs/resources/aws-marketplace.md, /docs/resources/aws-marketplace.md, /docs/resources/aws-marketplace.md

The documentation states that automatic model fallback tries the next model when the prior one "fails or is unavailable," but it does not specify the exact HTTP error codes or failure scenarios—such as 429 rate limits, 5xx server errors, or model timeouts—that trigger the automatic fallback chain.

Add an explicit section to the Model Fallback documentation listing which HTTP status codes (e.g., 429, 5xx) and failure modes (e.g., timeouts, connection errors) trigger automatic fallback versus returning an error directly to the client.

Where we looked: /docs/agent-api/model-fallback.md, /docs/agent-api/model-fallback, /docs/agent-api/model-fallback.md, /api-reference/agent-post.md, /docs/sdk/error-handling, /docs/sdk/error-handling.md

The Perplexity Search API (the `POST /search` web Search endpoint) accepts a `max_results` parameter with values from 1 to 20, and the default maximum is 10 results per search. None of the retrieved chunks mention pagination parameters, offset tokens, or any continuation mechanism, so the documentation does not clarify whether 20 is a hard cap or if additional results can be fetched via pagination.

Add explicit documentation on whether the Search API supports pagination (e.g., offset or page tokens) or if the max_results=20 limit is a hard cap per request.

Where we looked: /docs/search/quickstart.md, /docs/search/quickstart