Perplexity

Prompting Perplexity

Perplexity prompting is search prompting: split durable instructions from the concrete query, and route API work through the Agent API and Search API surfaces.

Which model to choose

Use Perplexity when current, cited information is the product. Start with the Agent API for essentially all new work: it runs web search, URL fetch, and tools behind tier presets (fast, low, medium, high, xhigh) and can route across multiple providers. The older Sonar chat-completions surface still works but is deprecated and supported only until September 27, 2026, so migrate to the Agent API. Use the Search API when you want ranked retrieval primitives rather than a generated answer.

Prompting in Perplexity

In the Perplexity GUI, phrase the prompt like a precise research query. Name the entities, time frame, geography, and comparison criteria. The same words that set your answer shape also influence what the system searches.

This does not mean Perplexity is always better than ChatGPT with web search. OpenAI’s web search tool also provides up-to-date answers with citations when enabled. Perplexity is the cleaner first stop when ranked search results, source filtering, and citation-forward research are the main deliverable. Its Search API returns ranked real-time web results with domain, language, region, and content-extraction controls, while the Agent API is designed for LLM-generated answers with citations.

Bad: “Tell me about AI pricing.”

Better: “Compare current published API pricing for OpenAI, Anthropic, Google, and Mistral for text generation as of August 2026. Cite the source for each input and output price.”

Prompting through the API

Through the API, keep durable behavior in instructions and put the concrete search-shaped question in input, then pick a preset (fast, low, medium, high, xhigh) or a specific model. The Agent API returns a typed output array (a message plus search_results items) rather than Sonar’s choices shape. Use request parameters for hard source, date, or domain constraints instead of prose when the surface supports them.

If you are still on Sonar chat completions, remember it is web-grounded but not omniscient, and that the surface is on a sunset path (2026-09-27). Ask it to cite and to state when search results are insufficient.

Current prompting guidance

• Make the query concrete enough to seed good search.
• Ask for citations on every factual claim.
• Tell the model to call out conflicts between sources.
• Prefer source filters and date controls over prose constraints when available.
• Split retrieval from synthesis when the research is high-stakes.

Examples

GUI prompt:

Find the current official API pricing for GPT-5.6 Sol, Claude Sonnet 5, Gemini 3.8
Flash, and Mistral Medium 3.5. Return input and output prices per million tokens,
with one citation per price.

API request skeleton (Agent API, POST https://api.perplexity.ai/v1/agent):

{
  "preset": "medium",
  "instructions": "Cite sources inline. If sources disagree, name the disagreement.",
  "input": "Compare the latest official model lists for OpenAI, Anthropic, Google, and Mistral as of August 2026."
}

Prompt upgrade example:

Stale pattern: “Research this topic and give me a broad overview.”

Better pattern: “Compare the current official documentation for these four providers, citing the exact page used for each model/version claim.”