Databricks + live news
Foundation models served next to your own tables. Relevant when the news you retrieve is joined against data you already hold.
| Vendor | Databricks, United States |
|---|---|
| Base URL | https:// |
| Model id | databricks-meta-llama-3-3-70b-instruct |
| Model credential | DATABRICKS_TOKEN |
| Tool format | OpenAI-compatible — the tool schema below is the same one every other OpenAI-shaped provider takes. Only the base URL and model name change. |
| News API key | none required |
| Vendor docs | https://docs.databricks.com/en/machine-learning/foundation-models/ |
Base URLs and model ids change often, and several vendors run separate mainland and international endpoints with non-interchangeable keys. Check the vendor docs above before copying. The part that does not change is the tool itself — our API has no key, no version negotiation and no SDK.
Working example
import json, os, requests
from openai import OpenAI
client = OpenAI(
api_key=os.environ["DATABRICKS_TOKEN"],
base_url="https://<workspace>.cloud.databricks.com/serving-endpoints",
)
TOOLS = [{
"type": "function",
"function": {
"name": "search_news",
"description": ("Search worldwide news published in the last 30 days. "
"Free, keyless. Use 2-5 keywords, not a question."),
"parameters": {
"type": "object",
"properties": {
"q": {"type": "string"},
"date": {"type": "string",
"enum": ["today", "yesterday", "24h", "7d", "30d"]},
"lang": {"type": "string"},
"size": {"type": "integer"},
},
"required": ["q"],
},
},
}]
IDENT = { # optional, but it is what keeps this API open
"agent_name": "news-agent",
"software": "dbrx",
"model": "databricks-meta-llama-3-3-70b-instruct",
}
msgs = [{"role": "user", "content": "What happened with grain exports this week?"}]
while True:
r = client.chat.completions.create(
model="databricks-meta-llama-3-3-70b-instruct", messages=msgs, tools=TOOLS)
msg = r.choices[0].message
msgs.append(msg)
if not msg.tool_calls:
print(msg.content)
break
for tc in msg.tool_calls:
args = json.loads(tc.function.arguments)
data = requests.get("https://freenewsapi.ai/v1/search",
params={**args, **IDENT}, timeout=20).json()
msgs.append({"role": "tool", "tool_call_id": tc.id,
"content": json.dumps(data)[:60000]})Worth knowing about Databricks
- The base URL contains your workspace host.
- External model endpoints proxy other vendors, so the same URL can front Claude or GPT with unchanged code.
Test the tool without the model
Before debugging a tool-calling loop, confirm the data side works. It needs no credentials, so this runs anywhere:
curl "https://freenewsapi.ai/v1/search?q=grain+exports&date=7d&size=5"If that returns results and your loop still fails, the problem is in the model's tool handling, not in the data.
Prompt it properly
The single most common failure is the model phrasing the query as a sentence. Search is
AND across every term, so q=what happened with grain exports matches nothing.
Put this in the system prompt:
Use search_news for anything about current events or recent developments.
Phrase the query as 2-5 keywords that would appear in a headline, never as a question.
Good: "grain export prices". Bad: "what is happening with grain exports".
Scan titles first. Only request full_text for articles you will actually cite.
If a search returns nothing, drop the least essential keyword and try once more.
No results means no coverage in this corpus \u2014 not that the event did not happen.More in system prompts.
Same format
Databricks uses the openai tool-calling shape. These take the same code with a different base URL and model name: Qwen, DeepSeek, MiniMax, Kimi, GLM.
