Image search
Image search lets a model find images on the web during a request. Gateway runs the search with Parallel and returns image URLs, titles, and source pages to the model. For text results, use web search.
Send a request with image search
Add {"type": "merge:image_search"} to tools. The model decides when to search, and each search can include up to three queries. Results are passed to the model as untrusted data.
It also works on the OpenAI- and Anthropic-compatible endpoints. On Anthropic Messages, name is optional:
Parameters
Set these in the tool’s parameters object.
Read citations and usage
Each image the model uses comes back as a url_citation annotation. url is the source page and image_url is the image. width and height are included when known. A citation is only added when the answer text contains the image or page URL.
Search counts appear in:
usage.server_tool_use.image_search_requestscounts searches, fast and advanced together.- OpenAI-compatible responses add
tool_usage.image_search.num_requests.
Combine with web search
Add both tools to the same request:
Each tool has its own parameters and is billed separately.
Pricing
Image search is billed per search, separately from tokens. Your organization’s standard margin applies.
Price doesn’t depend on the number of images returned. Failed searches aren’t billed.
Limitations
- Not available in batch inference or Fusion.
- Not available with zero data retention. Requests from those organizations are rejected.
- Results are links to third-party images. Make sure you have the rights to use them.
- The model’s search queries are checked against your organization’s DLP policy before they’re sent. A blocked query returns a tool error to the model and isn’t billed.
- Only one image search tool per request.