Skip to main content

Web Search

Level: Beginner · Category: API reference

Tags: Search · Grounding · Retrieval

The LLM API can search the public web in two ways: a search endpoint that returns results for your own code, and a web_search tool that some chat models use on their own. Both go through a SearXNG metasearch server run by Research Computing, so your search queries are sent to public search engines (your chat prompts are not). Keys and setup are on LLM API Access.

Search endpoint​

POST https://openai.rc.asu.edu/v1/search returns titles, URLs and snippets:

curl -s https://openai.rc.asu.edu/v1/search \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "Gaudi 2 HPU memory", "max_results": 3}'
{
  "id": "…",
  "object": "search",
  "results": [
    {"title": "Habana Gaudi2 AI Processor for Deep Learning …", "url": "https://www.intel.com/…", "snippet": "…", "date": null, "last_updated": null}
  ]
}
FieldMeaning
queryRequired. A string (or a list of strings, joined), up to 2,000 characters.
max_results1 to 20
search_domain_filterUp to 20 domains to keep, for example ["arxiv.org", "nature.com"]. Prefix a domain with - to exclude it.
time_rangeday, month or year
language, categories, engines, safesearch, pagenoPassed to SearXNG

GET /v1/search/tools lists the search backends your key can use. There is one today, searxng-search, so you don't need to name it. Searches are free and take a slot in the search server's own queue, separate from the models.

Some chat models have the web_search add-on (boon_details.web_search.active in /v1/models; glm-5-3 is one). The gateway gives the model a web_search tool and runs the searches it asks for, so the model can answer questions about recent events and cite sources. You call the model as usual:

from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY and OPENAI_BASE_URL
completion = client.chat.completions.create(
    model="glm-5-3",
    reasoning_effort="low",
    messages=[{"role": "user", "content": "Search the web: what is the latest released version of vLLM? One sentence, with the source."}],
)
print(completion.choices[0].message.content)
  • The model decides whether to search. Asking it to ("search the web", "check current sources") makes it more likely.
  • Search results are added to the prompt, so these requests use more prompt tokens and take longer.
  • The x-obleth-boons response header includes web_search when the add-on was active, and the header x-obleth-boons: off on your request turns it off (see Gateway add-ons).