Web Search Injection - MetrixLLM

Web Search Injection allows the gateway to automatically execute a web search, parse the results, and prepend the factual context to the system prompt before forwarding the request to the LLM. This happens entirely server-side. Your application does not need to handle search APIs or format the results.

Basic Usage

To trigger a web search, simply include the "web_search": true flag in the JSON body of your request to the gateway.

{
  "model": "gpt-4o",
  "messages": [{"role": "user", "content": "What is the stock price of Apple today?"}],
  "web_search": true
}

{
  "model": "claude-opus-4-5",
  "max_tokens": 1024,
  "messages": [{"role": "user", "content": "Latest news about SpaceX launch."}],
  "web_search": true
}

Advanced Options

You can fine-tune the search behavior by passing additional sub-parameters in your JSON payload:

Parameter	Type	Default	Description
`web_search`	`bool` \| `string`	`false`	Set to `true` to use MetrixLLM’s web search, or `"native"` to use the provider’s built-in search tool (if the provider supports it, e.g. OpenAI, Anthropic, Google).
`web_search_level`	`string`	`"medium"`	Determines the depth of the search. Options: `"low"`, `"medium"`, `"high"`.
`web_search_lang`	`string`	`"en"`	The language code for localized search results (e.g. `"fr"`, `"es"`).
`web_search_country`	`string`	`"us"`	The country code for geographically relevant results (e.g. `"uk"`, `"ca"`).
`web_search_max_uses`	`integer`	`undefined`	Maximum number of web searches to perform per request. When set, the gateway will stop searching after this limit is reached. Useful for controlling costs on complex multi-turn prompts.

Example: Localized search

curl https://gateway.metrixllm.com/openai/v1/chat/completions \
  -H "Authorization: Bearer mtx_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [{"role": "user", "content": "Dernières actualités de l'\''IA"}],
    "web_search": true,
    "web_search_lang": "fr",
    "web_search_country": "fr"
  }'

Example: Limiting search uses

curl https://gateway.metrixllm.com/openai/v1/chat/completions \
  -H "Authorization: Bearer mtx_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [{"role": "user", "content": "Compare the top 5 AI startups funding rounds this week"}],
    "web_search": true,
    "web_search_max_uses": 3
  }'

Example: Native web search (OpenAI)

When using "native" mode, the gateway delegates the web search to the provider’s built-in tool (e.g. OpenAI’s web_search_preview tool). This is useful when you want the provider to handle search context directly in its reasoning.

curl https://gateway.metrixllm.com/openai/v1/chat/completions \
  -H "Authorization: Bearer mtx_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [{"role": "user", "content": "What are the latest developments in quantum computing?"}],
    "web_search": "native"
  }'

Native web search is only supported by certain providers (OpenAI, Anthropic, Google). When using native mode, the provider’s own search tool handles query execution and result formatting. The web_search_level, web_search_lang, and web_search_country parameters are ignored in native mode since the provider controls search behavior.