What "Ask Medows" Actually Means

"Ask Medows" is the consult button on every patient card. It looks like a search box. It is not a search box.

The architecture behind it is what makes it different from typing the case into ChatGPT or UpToDate or OpenEvidence. Three components:

Context retrieval

When the doctor opens the AI consult, the patient's current state — vitals, recent labs, completed tasks, active medications, demographics, the last twelve hours of trend data — is automatically retrieved and embedded into the conversation. The AI does not see the doctor's question alone. It sees the question and the patient.

The word doing the work in that paragraph is trend. Sending the latest potassium is easy and nearly useless; a value without its history has to be interpreted against a population reference range, which is the least informative comparison available. Sending the last four values, with their timestamps, lets the model say something a textbook cannot: that this potassium is rising, that it is rising while the patient is already on a potassium-sparing diuretic, and that the rate matters more than the number.

What is deliberately not sent is anything that identifies the patient beyond what the clinical question needs. That boundary is written down in the privacy policy rather than left to a design instinct, because it is the kind of thing that erodes quietly.

Source-grounded generation

The AI's response is generated with reference to a structured corpus of clinical guidelines — AIIMS clinical protocols, ICMR documents, NICE guidance, drug references. Every claim in the response is attributable to a source. The citations are inline, with chip-style references the doctor can tap to see the underlying document.

The choice of retrieval over fine-tuning is deliberate and worth explaining, because it looks like an implementation detail and is not. A guideline changes. When the antibiotic policy is revised, a retrieval corpus is updated that afternoon and every answer from that moment cites the new document. A model with the old policy baked into its weights keeps confidently reciting last year's regimen, and there is no way to inspect it, no way to diff it, and no way to tell a clinician which version they are talking to.

Retrieval also makes the system auditable in the only way that matters clinically: you can ask where did that come from and get an answer that is a document rather than an assurance.

The interface lives in the patient's surface

The consult is not a separate tab. It is a panel that opens within the patient's card, with the patient's vitals visible above the conversation. The doctor doesn't switch context. The AI doesn't lose context.

That layout is also a safety mechanism, not just a convenience. If the model has picked up the wrong patient, or is reasoning from a stale set of vitals, the mismatch is visible in the same glance as the answer — the numbers it used are on screen, directly above the text it produced. An AI in a separate tab can be wrong about which patient it is discussing and there is nothing on the screen to contradict it.

What we mean by context-aware

The combination is what we mean by "context-aware AI." Not "an AI that knows everything," which is impossible and unhelpful. "An AI that knows this patient," which is achievable and useful.

The published frameworks for this — retrieval-augmented generation, RAG, grounding via document retrieval — have been mature for a couple of years now. What's new in Medows is the clinical-workflow layer. The doctor doesn't think about retrieval. They think about the patient. The retrieval happens because the patient's card is the starting point of the conversation.

The failure modes we design against

Being specific about how this breaks is more useful than describing how it works.

Retrieval comes back with nothing relevant. The dangerous response is a fluent answer generated from the model's general knowledge, dressed in the same confident register as a sourced one. The correct response is to say the corpus does not cover this and answer without pretending otherwise — visibly ungrounded, so the doctor knows to weight it differently.

The context is stale. A result that landed two minutes ago and has not been pulled in makes the answer wrong in a way that is very hard to see. The context has to be fetched at question time, not at page load.

A confidently wrong number. Doses, rates, and intervals are where a plausible-sounding error does the most damage, and they are exactly what a language model is worst at. These get the citation treatment hardest: a dose without a source beside it should read as a claim, not an instruction.

Over-trust after a good run. The most likely long-run failure is not a single bad answer. It is fifty good ones, followed by a doctor who has stopped reading the citations. This is why the sources stay inline and tappable rather than collapsing into a footnote after the novelty wears off — the friction is doing something.

Three things "Ask Medows" is not

• Not a diagnosis tool. The output is decision support, not a decision.
• Not a substitute for the senior consultant. The output sits next to the patient context it used, so the consultant can verify.
• Not a clinical search engine. The starting point is the patient, not the question.

The point of the architecture is to remove the cognitive overhead of asking the question. Restate the case. Look up the guideline. Verify the dose. Check the interaction. Five steps become one — for the AI part. The doctor's clinical judgment is still the doctor's clinical judgment.

That division is the whole design. The machine does retrieval, arithmetic, and recall, which it is reliable at. The doctor does judgement, which it is not. Any product that blurs that line is making a promise it cannot keep, and the first person to find out will be a patient.