Perplexity Search

Perplexity Sonar models generate web-search-based responses. The gateway lets you control search behavior in two ways:

1. Model slug suffix — perplexity/sonar:high, perplexity/sonar-pro:pro
2. Body field — { "web_search_options": { "search_context_size": "high", "search_type": "pro" } }

If both are provided, the body field wins.

Search Context Size

Sonar runs each request with a search_context_size. This affects search depth and how much outside context the model uses.

Size	Suffix alias	Description
low	:low	Default, fastest and cheapest
medium	:medium, :med	Balanced depth/cost
high	:high	Maximum search depth, for research

Pricing can vary by model and mode. Confirm the current price before production use.

Pro Search (sonar-pro)

sonar-pro offers a deeper search mode for more complex questions. Trigger it with a suffix or body field:

curl https://llmtr.com/v1/chat/completions \
  -H "Authorization: Bearer sk_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "perplexity/sonar-pro:pro",
    "messages": [
      {"role": "user", "content": "What was Turkey's annual inflation in 2025, with sources?"}
    ]
  }'

Pro Search is only supported on sonar-pro; using :pro on another model returns 400 unsupported_capability.

Deep Research (sonar-deep-research)

Deep Research is meant for broader research and source gathering. Cost and latency can differ noticeably from standard Sonar calls.

Unknown suffix

Unknown suffixes such as perplexity/sonar:turbo return 400 invalid_request_error — they are not silently swallowed, so a typo in the slug never sneaks into your invoice unnoticed.

Embeddings

Perplexity also offers pplx-embed-v1-* and pplx-embed-context-v1-* embedding models. Standard embeddings call /v1/embeddings; contextualized embeddings use the same gateway endpoint with a nested array input:

curl https://llmtr.com/v1/embeddings \
  -H "Authorization: Bearer sk_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "perplexity/pplx-embed-context-v1-0.6b",
    "input": [
      ["Document 1, chunk 1.", "Document 1, chunk 2."],
      ["Document 2, single chunk."]
    ]
  }'

In contextualized embeddings, the outer array represents documents and the inner array represents chunks for each document.