Under the hood
How your agent works, in plain terms.
You do not need a technical background to build on Ostify. This page is for when you, your clinical safety officer or your IT team want to know what is happening underneath.
Where answers come from
Your agent looks things up. It does not answer from memory.
A general AI chatbot answers from everything it has ever read, which is why it can sound confident and be wrong. An Ostify agent works more like a colleague who checks your leaflet before replying. The technical name is retrieval-augmented generation, or RAG.
- Your content is split into short passages
Each leaflet or guideline is broken into pieces a few paragraphs long. You read them, correct them and approve them. The agent can only use the ones you approve.
- It finds the passages that fit the question
The search goes by meaning, not exact words, so “will it hurt?” can find your passage on pain relief. It only ever looks through your agent’s own content.
- No good match means no answer
If nothing in your content is close enough to the question, the agent gives your out-of-scope message. It does not try to fill the gap.
- The answer is written from those passages
The AI model is handed the matching passages and told to use those and nothing else. Every answer shows which passage it came from, so you can check it.
Gating
A question passes a series of gates before anything is answered.
A gate is a checkpoint. A question that does not pass is stopped there and given a fixed, safe message. The six guardrails list what each one looks for.
- Urgent questions never reach the AI
Emergencies and requests for a clinical decision are caught by the first gates. The patient gets your signposting message straight away, and the AI model is never asked.
- Checked on the way in and the way out
The question is screened for harmful content before an answer is written. The finished answer is screened again before the patient sees it.
- Patients only see the reviewed version
Your patient link uses the version we reviewed. If you change your content or key settings, the link pauses until the new version has been reviewed.
The AI models
Where the AI runs. It never learns from your content.
The model is the part that reads a question and writes a reply. Ours run in Ostify’s own environment on Microsoft Azure, not on a public chatbot service. What happens to the data itself is on the trust page. How the platform is protected is on the cybersecurity page.
- Models that answer patients
- Run in the UKA patient’s question is read and answered by models hosted in the UK.
- Models that test and draft
- Run in the EUOsti-gen, Osti-judge and Osteocite Assist use a different model from the one that answers patients, so an answer is never marked by the model that wrote it. They see your test questions and your settings, not patient conversations.
- Training
- Never used to train AIYour content, the questions your agent is asked and the answers it gives are not used to train or improve any AI model.