perplexity

Package

@ai-sdk/perplexity

Factories

AIProviders.perplexity

Capabilities

• Language
• Embedding
• Decision

Perplexity 5 routing

languageModel(...) and the callable provider use /v1/agent. Choose chatModel(...) or sonarModel(...) for legacy Sonar Chat Completions. embeddingModel(...) uses /v1/embeddings.

Native decisions

let perplexity = try AIProviders.perplexity()
let result = try await AI.experimentalDecide(
    model: try perplexity.decisionModel("pplx-judge"),
    state: .text("The package arrived cracked."),
    questions: ["damaged": .boolean(criteria: ["true": "Package is damaged"])]
)

The /v1/decisions endpoint accepts 1–128 named Choice, Score and Boolean questions. Choice supports 1–255 options; Score requires 1–10 non-null levels. Instructions may be omitted when the rubric defines the judgment. Box is unsupported. State preserves ordered text/JSON and up to eight inline PNG, JPEG or WebP images, including images nested in JSON. The facade resolves URLs before adapter execution.

Answers retain full precision without rounding allowance. Confidence and score legends live under providerMetadata.perplexity; response metadata keeps x-request-id, resolved model, headers and raw body. maxOutputTokens, reasoning and Perplexity provider options produce unsupported warnings.

Notes

Perplexity Agent language models, legacy Sonar chat, and quantized embeddings mirror upstream Perplexity 5.0.20: PERPLEXITY_API_KEY, custom base URL and headers, versioned user-agent suffix, callable/language Agent surface plus chat/sonar compatibility and embedding/textEmbedding aliases, /v1/agent, /chat/completions, and /v1/embeddings, text/image/PDF message conversion, JSON schema response format, providerOptions.perplexity passthrough including quantized embedding dimensions and signed/binary base64 decoding with a 512-value preflight limit, citations/images/cost metadata, optional and null streaming delta content, reasoning-token separation from completion tokens with corrected total accounting, Perplexity error schema, and warnings for unsupported topK/stop/seed/custom reasoning while treating provider-default reasoning as a no-op. Perplexity 5.0.1 accepts explicit null values for optional Agent SSE fields without failing the stream, including null content indexes that retain the base text-part ID. The 5.0.0 migration makes languageModel and the callable provider use the Agent API; use chatModel or sonarModel for the legacy Sonar Chat Completions API. ProviderV4 type names, ESM-only packaging, Node 22 engines, optional JS headers config, and workflow serialization helpers are JS-only upstream concerns. The default X-Pplx-Integration header is vercel-ai-sdk; explicit headers override it case-insensitively. The 2026-10-11 pass: Common max effort maps to xhigh; unknown effort warns and is omitted. Embeddings inherit shared dimensions unless a provider override is supplied. The 2026-10-23 pass: Native decisionModel calls /v1/decisions with named Choice/Score/Boolean questions, up to eight inline images, full-precision answers, confidence/score legends, and unsupported-setting warnings.

Factory argument requirements are defined by the public Swift factory signatures. Use Public symbols when you need the exact initializer or factory declaration.

Return to the provider matrix.