Grounding LLM answers in search results

1. 01Search the question

   Send the user's question, or a query the model writes from it, with the country and language of the user. Google's order is a useful relevance signal to keep.

   GET /v1/search?engine=google_search&q=albert+einstein&country=us&page=1 · query

   {
     "q": "albert einstein",
     "language": "en",
     "country": "us",
     "page": 1
   }

2. 02Keep the top results as sources

   Each result has a title, URL, snippet, and a date when Google shows one. Number them in the prompt and ask the model to cite by number.

   GET /v1/search?engine=google_search&q=albert+einstein&country=us&page=1 · results.0

   {
     "position": 1,
     "title": "Albert Einstein",
     "url": "https://en.wikipedia.org/wiki/Albert_Einstein",
     "source": "Wikipedia",
     "displayed_url": "https://en.wikipedia.org › wiki › Albert_Einstein",
     "date": null,
     "snippet": "Albert Einstein (14 March 1879 – 18 April 1955) was a German-born theoretical physicist ...",
     "sitelinks": []
   }

3. 03Add the knowledge graph when there is one

   For people, places, and organisations, knowledge_graph carries a short description with its source and key facts. It is often the most direct answer on the page.

   GET /v1/search?engine=google_search&q=albert+einstein&country=us&page=1 · knowledge_graph

   {
     "title": "Albert Einstein",
     "type": "Theoretical physicist",
     "description": "Albert Einstein was a German-born theoretical physicist best known for developing the theory of relativity.",
     "description_source": {
       "title": "Wikipedia",
       "url": "https://en.wikipedia.org/wiki/Albert_Einstein"
     },
     "facts": {
       "Born": "March 14, 1879, Ulm, Germany",
       "Height": "5′ 9″"
     }
   }

# Ground a model's answer in current Google results: search, keep the top
# results with their snippets, and pass them as numbered sources.
import os, requests

API = "https://api.clair.im"
HEADERS = {"Authorization": f"Bearer {os.environ['CLAIR_API_KEY']}"}

def google(q, country="us", language="en"):
    r = requests.get(f"{API}/v1/search", headers=HEADERS, timeout=60, params={
        "engine": "google_search", "q": q, "country": country, "language": language})
    r.raise_for_status()
    return r.json()

def sources_for(question, k=6):
    body = google(question)
    lines = []
    if body.get("knowledge_graph"):
        kg = body["knowledge_graph"]
        lines.append(f"[0] {kg['title']} ({kg['type']}): {kg['description']}")
    for r in body["results"][:k]:
        date = f" ({r['date']})" if r["date"] else ""
        lines.append(f"[{r['position']}] {r['title']}{date}\n{r['url']}\n{r['snippet']}")
    return "\n\n".join(lines)

question = "What changed in the EU AI Act this year?"
prompt = (
    "Answer using only the sources below, and cite them by number. "
    "Say so if they do not answer the question.\n\n"
    f"Sources:\n{sources_for(question)}\n\nQuestion: {question}"
)
print(prompt)  # send to the model of your choice

Set CLAIR_API_KEY to a key subscribed to the Google Search API. Each call is one request against that API's monthly quota.

• Snippets are short. For questions that need the page's content, fetch the top URLs yourself and pass the text.
• A live search takes seconds, and more at the slow end. Show a progress state, or search while the model plans.
• Clair returns organic results without ads or Google's AI Overview. That is usually what you want as sources.
• Google's results reflect country and language. Pass the user's, or the answers will cite sources for the wrong market.