LLM-side Web Search tool

LLM-side Web Search lets AI models retrieve up-to-date information from the Web without requiring a separate search integration. Search is performed by the model provider, and relevant results are incorporated into the response. This is useful for questions that depend on recent, specific or external information beyond the model’s built-in knowledge.

The LLM Mesh supports natively the server-side Web Search tools available from several LLM providers through an unified API.

Configuration

A DSS administrator must:

• enable Web Search on the LLM connection configuration page

• make sure it is supported/configured at the model level

Usage

If the selected LLM supports it, you can use the Web Search tool from:

• Prompt Studios, Prompt Recipes: configure it in the completion settings

• Visual Agents: enable the Web Search tool

Web Search is also available through the REST and Python APIs as a web_search tool type:

completion = llm.new_completion()
completion.settings["tools"] = [
    {
        "type": "web_search", # mandatory
        "allowedDomains": ["google.com"], # optional, mutually exclusive with blockedDomains
        "blockedDomains": ["google.com"], # optional, mutually exclusive with allowedDomains
        "userLocation": { # optional
            "type": "approximate",
            "country": "FR",
            "region": "Ile-de-France",
            "city": "Paris",
            "timezone": "Europe/Paris",
        },
        "maxUses": 3, # optional, only supported by Anthropic models
        "failIfUnsupported": True # optional, default: True
    }
]

Output

When a Web search call is made and returns results, those are returned as sources by the API and displayed in visual components, as well as included in the trace.

Limitations

The Web Search tool is only supported for the following connections:

• OpenAI

• AzureOpenAI

• Microsoft Foundry

• Anthropic

The Web Search tool is only supported for the following models:

• OpenAI models using the Responses API

• Anthropic models using the Messages API