Drug-Safety Agent on Databricks · Databricks · Foundation Model APIs
Foundation Model APIs
Foundation Model APIs host open models so you do not run them. They speak the same request format as the OpenAI API, so the agent's model port connects to them with a base URL and a model name.
Which model. A strong open model such as gpt-oss-120b suits this job, which is tool calling with short structured answers. For choices across vendors see the site's which model page.
What is true today. An adapter for the Databricks endpoint is written. The 20-pair evaluation ran on the same open model through another host. Running the identical evaluation on the Databricks endpoint is the step that would make the "built on Databricks AI" claim fully evidenced.
The real code
Cut from the repository each time this site is built, so it always matches what runs. Each header names the file and the line it was cut from.
src/drugsafety/agent.py:157The Databricks adapter: an OpenAI-style chat call to a serving endpointclass DatabricksModel:
"""Foundation Model APIs (OpenAI-compatible chat). Untested while the workspace refuses model queries."""
def __init__(self, endpoint: str = "databricks-gpt-oss-120b", profile: str | None = None):
import os
import subprocess