Internet Search

The internet_search tool is a built-in tool automatically available on all /v1/chat/completions requests. It uses Perplexity AI Sonar under the hood to search the live web and return a synthesized answer with citations. You don't configure it — the model invokes it automatically when the prompt requires real-time or current information.

How It Works

1. You send a normal chat request with a question that needs current data

2. The LLM detects it needs fresh web data and calls internet_search internally

3. The tool queries Perplexity Sonar (temperature: 0.1 for accuracy) with your query

4. Perplexity returns a synthesized answer + citation URLs

5. The model wraps the answer into a standard chat completion response

Streaming Behavior

Unlike image generation (which delivers the result at the end), internet search integrates naturally with streaming. The stream shows:

1. A tool_calls chunk announcing internet_search was invoked

2. Immediately followed by content chunks streaming the answer word-by-word

3. A final [DONE] chunk

4. A trailing usage data chunk (if stream_options.include_usage: true)

data: {"delta":{"tool_calls":[{"function":{"name":"internet_search",...}}]}}
data: {"delta":{"content":"As"}}
data: {"delta":{"content":" of"}}
...
data: [DONE]
data: {"usage":{...}}

Use Cases & Examples

1. Breaking News / Current Events

{
  "model": "google/gemini-2.5-flash",
  "messages": [
    { "role": "user", "content": "What are the latest AI news today?" }
  ]
}

2. Live Price / Market Data

{
  "model": "google/gemini-2.5-flash",
  "messages": [
    { "role": "user", "content": "What is the current price of Bitcoin?" }
  ]
}

3. Combining Search + Reasoning

{
  "model": "google/gemini-2.5-flash",
  "messages": [
    {
      "role": "user",
      "content": "Search for the latest GPT-5 benchmarks and summarize how it compares to Claude Sonnet"
    }
  ]
}

4. Technical / Developer Queries

{
  "model": "google/gemini-2.5-flash",
  "messages": [
    { "role": "user", "content": "What are the breaking changes in Bun 2.0?" }
  ]
}

Tips & Best Practices

Tip	Details
Ask naturally	No special syntax needed — the model detects when to search automatically
Be specific	"Latest React 19 release notes" works better than "React news"
Date context helps	Adding "as of today" or "current" reinforces real-time lookup
Streaming for UX	Use stream: true — answers start flowing within ~1s while search runs
Cost vs knowledge cutoff	For questions within the model's training data, search won't trigger (saves cost)
Citations in output	The model includes source links inline — these are real URLs from Perplexity

---

Limitations

• Powered by perplexity/sonar — accuracy depends on Perplexity's index freshness

• No control over numResults from the request body — handled internally (default 5 citations)

• Does not return raw search results, only the synthesized Perplexity answer

• Adds latency (~1–3s) compared to non-search requests

• temperature is fixed at 0.1 internally for factual accuracy — not overridable