Top Health AI · Document-Grounded Chatbot
Patients Ask. It Answers Only From Your Pamphlet
Patients text questions at 9pm that the front desk already answered a hundred times, from a pamphlet that never left the exam room. We turned that pamphlet into a chatbot, reachable by text or on the web, that answers only from what the provider uploaded, and says so plainly when it can’t.
The rest get the safe fallback, by design rather than by accident
Time the front desk gets back for the patients who need a person
Three in ten patients ask when the office is closed
Routine questions stop competing with scheduling and billing
The problem
Patient Education Pamphlets Don’t Answer Questions After 9pm.
Every practice hands out physician-approved pamphlets on aftercare, medication, and procedures. They’re accurate, but static. The moment a patient has a follow-up question at home, the pamphlet can’t answer it, so the call goes to a front desk that’s already fielding the same handful of questions all day.
A generic AI chatbot looks like a fix until it invents an answer the pamphlet never gave. In healthcare, that’s not a bad user experience, it’s a liability. Most chatbot options on the market are either a basic FAQ box with no real understanding, or a full EMR integration far beyond what a practice needs to answer “can I shower with this bandage on.”
The brief wasn’t “build a chatbot.” It was “build one that is structurally incapable of guessing about a patient’s health.”
Kodexo Labs project team
What we built
A Pamphlet That Talks Back, and Only About What’s Actually in It.
A provider uploads their approved pamphlets once. From then on, patients ask questions in plain English by text or on the web, and every answer is pulled from that provider’s own documents, never generated from scratch.
01 · Patient Q&A
Ask by Text or on the Web, Get an Answer From the Pamphlet.
A patient texts the provider’s number, or asks on the web and watches the answer stream in word by word. No app to download, no keywords to guess. The system reads the question, finds the matching section of the pamphlet, and replies in a sentence or two.
02 · Safety controls
If It’s Not in the Pamphlet, It Says So.
Every question is checked against the approved content before an answer is written. Nothing outside it gets a guess; the patient is pointed back to their provider instead. More on this below.
03 · Provider & admin tools
Every Provider Runs on Their Own Private Set of Pamphlets.
An admin brings a new provider on with an emailed invite, uploads that provider’s pamphlets, and the provider gets a dashboard showing every patient conversation, so nothing happens out of view.
Built so it can never guess
The Chatbot Doesn’t Know Anything the Provider Didn’t Hand It.
Most chatbots answer from everything a language model has ever read, which is exactly the wrong instinct for medical information. This one is wired the other way: it can only draw on the pamphlets a provider has uploaded and approved. Ask it anything outside that, and it isn’t allowed to improvise.
When a question falls outside the pamphlet, the system says so in one consistent line, every time, with no partial guesses, no “it depends” filled in from elsewhere; “I don’t have information about that. Please consult your healthcare provider.”
Strict content boundary
Answers can only be built from the uploaded pamphlets, with no fallback to general knowledge.
Out-of-scope detection
Every question is checked against the pamphlet content before a response is written, not after.
One consistent fallback
The same safe message every time a question falls outside scope, never a partial or invented answer.
The solution
Four Parts, and Not One of Them Guesses.
The build is deliberately small. Four moving parts, each one chosen so a patient can get a fast answer at home without anyone inventing medical advice along the way.
Upload Once
A provider hands over their approved PDF pamphlets. The system reads them, including scanned ones, splits them into small passages, and indexes every passage so it can be found by meaning later. No re-typing, no tagging, no keyword lists.
Ask in Plain English
A patient writes the way they would speak. The system finds the closest matching passages in that provider’s own pamphlets and writes a short answer, one or two sentences, from those passages only.
Refuse When Unsure
Every question is checked against the pamphlet content before a reply is written. If nothing matches closely enough, the patient gets the same safe line every time and is pointed back to their provider.
Available All Day
It runs in the cloud, so patients reach it from home, the waiting room, or a phone at 9pm. Each provider’s documents and conversations stay separate from every other provider’s.
How it works
Three Flows, From a Doctor’s Pamphlet to a Patient’s Phone.
No jargon, just the path a question takes from a patient’s thumb to a safe, sourced answer.
System architecture
The Whole System on One Page.
The plain-language walkthrough is above. This is the same system named properly: clients, the service layer, the ingestion and query pipelines, and the two data stores.
Under the hood
The Tools We Used, and What Each One Does.
Pamphlets are chunked, embedded, and stored per provider; a retrieval step pulls only the closest matches before the model is allowed to answer.
Production stack
Provider-isolated data
Each provider’s pamphlets and chat history are kept separate from every other provider’s.
Encrypted onboarding
Provider invitations and password setup run over encrypted links, end to end.
24/7 cloud access
Patients reach it from home, the waiting room, or a mobile device, any time.
Full conversation record
Every exchange is saved against the patient and provider, for the provider to review.
The outcome
The Pamphlet Now Answers at 9pm, and the Front Desk Gets Its Afternoon Back.
Nothing about the clinical content changed — every answer is still the provider’s own approved text. What changed is when a patient can reach it, and how much staff time gets spent repeating it.
Patients per provider at once
Concurrent sessions handled without slowing down, so a busy practice never queues behind itself.
From PDF upload to searchable
A new provider’s pamphlets are live in about a business day and a half, extraction and indexing included.
Availability with no staff on shift
Reachable by text or on the web from home, the waiting room, or a phone, with no one rostered to answer.
Staff answer the hard questions
The repetitive, pamphlet-covered questions go to the chatbot, which leaves clinical staff available for the calls that actually need a person.
No answer the provider didn’t approve
Because a reply can only be built from uploaded content, there is no path for the system to invent medical advice — it says it doesn’t know instead.
Every provider stays separate
Pamphlets are indexed per provider, so one practice’s content never surfaces in another’s answers, and onboarding a new one adds no work to the existing ones.
Frequently asked
Common Questions About a Chatbot Like This
Where do the answers come from?
Only from the PDF pamphlets that provider uploaded and approved. The system has no other source to draw on, so it cannot offer general medical opinion.
What happens when a question is not covered?
The patient gets one consistent line: “I don’t have information about that. Please consult your healthcare provider.” Never a partial answer, never a guess.
Can patients use it without an app?
Yes. They text the provider’s number, or ask on the practice’s web page and watch the answer stream in. Nothing to download.
Can one provider see another provider’s content?
No. Pamphlets and chat history are stored against a single provider, and a question only ever searches that provider’s own documents.
Is it HIPAA compliant?
This first version runs on SOC2-compliant infrastructure. Full HIPAA-grade configuration, including agreements with every third-party service, is planned for the next phase.
Does the provider see what patients asked?
Every exchange is saved and shown in the provider’s dashboard, by patient, in full, so nothing happens out of view.
How do new pamphlets get added?
A provider sends the new PDF, it is uploaded and indexed, and it is answerable from that moment. No rebuild, no retraining.
What is not in this first version?
English only, no emergency detection or 911 routing, and no deeper reporting. Those are scoped for the next phase.
Have Patient Content That Should Be Talking Back?
30 minutes, and you’ll know whether a document-grounded assistant fits your pamphlets.
Related work
