Drug-Safety Agent on Databricks · End to end · Direct reads
Direct reads: counts and label, no model
If someone only wants the report counts, or only wants to know whether the label mentions an effect, a model adds nothing but cost and risk. These two requests are answered straight from the tools.
Unknown pairs. If the drug or event is not in the data, the reply lists what the data does hold. It does not guess and it does not return zeros, because a zero would look like a real finding.
Worked example. Asked for the counts, the reply is: lisinopril and ANGIOEDEMA: 7,015 reports with both; 309,818 for the drug without the event; 37,265 for other drugs with the event; 20,700,061 for other drugs without it (openFDA, fetched 2026-11-19).
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/entrypoint.py:44_direct(): one tool call, one fixed sentenceasync def _direct(r: Route, tools, who: Identity) -> str:
seen: list[dict] = []
facade = PairTools(tools, seen)
name = "get_report_counts" if r.intent == "counts_only" else "get_label_mentions"
key = IdempotencyKey(run_id=RunId(f"run_direct_{who.customer_id}"), step=0, iteration=0)
got = await facade.call(name, {"drug": r.drug, "event": r.event}, who, key)
out: dict[str, Any] = dict(got.structured) if isinstance(got.structured, dict) else {}
if "error" in out:
drugs, events = ", ".join(out["drugs"]), ", ".join(out["events"])
return f"The data does not hold that pair. It holds drugs: {drugs}; events: {events}."
if r.intent == "counts_only":
return (
f"{out['drug']} and {out['event']}: {out['a']} reports with both; {out['b']} for the drug without the event; "
f"{out['c']} for other drugs with the event; {out['d']} for other drugs without it "
f"(openFDA, fetched {out['fetched']})."
)
where = "yes" if out["on_label"] else "no"
return f"Does the {out['drug']} label mention {out['event'].lower()}? {where}. Label version {out['label_version']}."