Official health answers, while the line is busy.

Where to get care, how to book, where to get vaccinated. An agent answers those questions at any hour, in French or English, and cites the official page behind every answer. The nurse keeps the calls that need a nurse.

Try Agent Santé Talk to us

voice-question.webm

Citizen0:00I've had gastro since yesterday. What can I do?

Agent Santé0:06Québec.ca advises drinking more fluids than usual. An oral rehydration solution helps replace what you lose.

Agent Santé0:15I can't assess your situation. For that, call Info-Santé at 811. In an emergency, call 911.

Agent Santé0:31Here are three pharmacies near H2T, each linked to its listing in the government directory.

personal_info_checkPassed. Nothing personal in the message, so it goes on to the agents.

sourceQuébec.ca, quoted and linked.

medical_advice_checkPassed. No diagnosis, no treatment, no dose. 811 and 911 named.

nearby[]Three, from the government directory of health and social services.

The problem

The answer exists. The queue is in front of it.

Wait times at Info-Santé 811 reached up to two hours on average this fall, before the winter peak. Many of the calls in that queue ask where to get care, how to book an appointment or where to get vaccinated. Those are legitimate questions, and the answers are already published on official sites. They are just hard to find.

Every minute a nurse spends reading a web page aloud is a minute less for the caller who needs a nurse. An agent can take the first kind of question, at any hour, and send the second kind back to the people who can assess it.

The agent does not replace the nurse. It clears the line in front of one.

The workflow

Question in. Cited answer out.

Five stages. Two of them are checks the model does not get to skip.

InA question, typed or spokenFrench or English. No account, no name.

Transcribe
A voice message is transcribed in Montréal by rakeaudio. Only its text goes further. The recording is not kept.
rakeaudio-asrMontréal
Screen
decision-machine-1 looks for personal information before any agent sees the message. If it finds some, the message stops here.
decision-machine-1blocked, not forwarded
Route
An intake agent reads the question and hands it to a specialized agent. Gastroenteritis, respiratory viruses, vaccination, seniors, mental health and more.
intake agentone agent per topic
Answer from the source
The specialized agent searches the official sources it is allowed to read, quotes the page and links back to it.
Knowledge searchlink to the page
Check
decision-machine-1 reads the answer for personal medical advice. If it finds some, the answer is withdrawn and rewritten. After three withdrawals, the conversation stops.
withdrawn, rewritten3 strikes

OutA cited answerOr a referral to 811 or 911 when only a person can help.

Guardrails

Not left to the model's good behaviour.

A health agent has to earn trust. The rules that matter run as separate checks on every exchange, outside the language model that writes the answer.

Diagnosis

What people assume The prompt says "do not diagnose," and the model mostly listens.

What happens A second model checks every answer before it is shown. Personal medical advice is withdrawn and the agent must rewrite. The counter of withdrawn answers is public.
Personal information

What people assume Whatever the citizen types goes straight to a model provider.

What happens The message is screened in Montréal first. A message with personal information is blocked and never reaches the agents or the language models.
Where it came from

What people assume A fluent paragraph, and you take its word for it.

What happens Every answer quotes an official page and links to it. Québec.ca, Santé Québec, the INSPQ, the ministry of Health. Anyone can check in one click.
Booking

What people assume The agent books the appointment for you.

What happens It links to the Clic Santé results page and stops there. It does not read availabilities and it never books. That boundary is a choice, and the same foundation can go further when a health network asks.

Where it answers

Wherever the questions already arrive.

The agent is the same on every channel. Run gives it the doors.

A public website

Type or speak a question. No account, nothing to upload, and the conversation lives in the browser's memory only.

A phone line

Run assigns US and Canadian numbers to agents. The agent answers, and transfers the caller to a person when one is needed.

Your own product

Start runs over HTTP and stream each step as events. A patient portal or a clinic app can carry the same agent under its own name.

Where the data goes

Said plainly, because it has to be.

This is how Agent Santé runs today. Each of these is a setting, not a rewrite.

Stays in Montréal

The personal information check on every message, the medical advice check on every answer, and the transcription of every voice message.

Goes to the model servers

The text of a message the filter accepted, with the verified conversation so far. In Canada and the US. Never a blocked message, never a recording.

Kept nowhere

Server logs never hold the text of messages or the sound of voice messages. Exchanges are erased about 10 minutes after the conversation ends.

In production

Agent Santé agentsante.ca

This one is already answering.

Agent Santé is free for every Quebecer, in French and English, at any hour. CloudRaker built it as a volunteer project on Run, and its transparency page publishes the real counters, live.

Conversations: 151
Verified answers: 227
Messages blocked for personal information: 8
Answers withdrawn for medical advice: 29

Launch-day figures. The live counts are on the transparency page.

No account, no name. Personal information is blocked in Montréal before any agent sees it.
Conversations are erased about 10 minutes after they end.
The site code and the agent instructions are public under Apache 2.0.

Read the launch note See the live counters

Agent Santé in 30 seconds (in French).

What it runs on

The building blocks under this one.

One platform for the agents, two CloudRaker models for the checks, and the open code to read how it fits together.

CloudRaker Run

The intake agent and every specialized agent, their knowledge folders, and a record of every run with what it cost.

See Run

decision-machine-1

CloudRaker's model for decisions, hosted in Montréal. It runs both checks, on the way in and on the way out.

See the model

rakeaudio

CloudRaker's transcription model, in Montréal. A voice question becomes text before anything else reads it.

See the model

The open code

The site and the agent instructions, public under Apache 2.0. Read them, verify them, start from them.

See the agents on GitHub

Related workflows

Same agents. Different front doors.

Anywhere people wait in a queue for an answer that is already written down, or for a person who needs the paperwork done first.

Medical scribeWhen the call does reach a clinician, the conversation becomes a structured, cited note.Open

Clinical intakeForms, faxes and referral letters into one structured record, with field-level review.

Referrals and triageParsed, routed by specialty and urgency, and acknowledged.

Start here