Promeing ConceptScenario set in Geneva
Clinical Knowledge Intelligence
IndustryHealthcare
Visualisation
- Status
- Promeing Concept
- Scenario
- Geneva
- Industry
- Healthcare
- Capabilities
- AI InfrastructureAI AgentsAI Products
- Sheet
- 05 / 06
Private retrieval over a hospital’s own guidelines and protocols, running inside its own infrastructure, with paragraph-level citations and a refusal whenever sources conflict.
A Promeing Concept: a studio study of a system we would build, set in a realistic scenario. It describes no client, no deployment and no measured result.
Problem
Clinicians search guidelines, protocols and formularies spread across intranets, shared drives and PDFs. The right document exists, but finding the current version of the right paragraph takes time that a ward does not have, and an outdated copy looks exactly like a current one.
Context
The scenario: a hospital in Geneva with departmental protocols, hospital-wide guidelines and a formulary, each revised on its own schedule and published in French and English. Patient data must never leave the hospital, any tool must be explainable to clinical governance, and nothing may present itself as making a clinical decision.
System
Private retrieval running entirely inside the hospital’s own infrastructure. Documents are indexed with their version and status, so withdrawn versions are never cited. Every answer quotes the exact paragraph and document version it came from, and the system declines to answer when sources are missing or disagree, naming the conflict instead. It is decision support for clinicians, never diagnosis or decision-making. Every query is logged for clinical governance, and the system is evaluated with clinicians against questions they define.
AI architecture
Documents are versioned and indexed inside the hospital. A grounded answer step can only use retrieved paragraphs, a conflict check decides whether to answer at all, and every query lands in a governance log.
AI architecture of Clinical Knowledge Intelligence
Layers
- Interface: Clinician answer view, Governance review
- Agents: Grounded answer step, Conflict and refusal check
- Retrieval: Paragraph-level retrieval
- Models: On-premise language model
- Data: Versioned document library
- Infrastructure: Hospital infrastructure, Query and answer log
Connections
- Versioned document library → Paragraph-level retrieval
- Paragraph-level retrieval → Grounded answer step
- On-premise language model → Grounded answer step
- Grounded answer step → Conflict and refusal check
- Conflict and refusal check → Clinician answer view
- Conflict and refusal check → Query and answer log
- Query and answer log → Governance review
- Hospital infrastructure → On-premise language model
Interface
The answer view: a question, a short answer with paragraph-level citations, the sources with their versions, and the state that matters most: sources disagree, not answering.
Which version of the anticoagulation protocol applies on the cardiology ward, and what does it say about bridging before surgery?
The ward follows protocol AC-14, version 6, in force since 2026-03-01.1 It states that bridging is decided by the treating team for each patient, using the risk table in section 4.2
Which of the two agents is preferred after this procedure?
Flagged to the pharmacy team for review
Logged for clinical governanceQ-20261002-1187
-
1Ward handbook, cardiology
v3 · §2.1 ¶1
Anticoagulation on this ward follows protocol AC-14 in its current version.
-
2Protocol AC-14
v6 · §4.2 ¶3
Bridging is decided by the treating team for each patient, using the risk table in section 4.
- A: An answer that can be checked: each sentence points to the paragraph it came from.
- B: Sources carry their document, version and paragraph, so an outdated copy is never mistaken for a current one.
- C: When sources conflict, the system says so and does not answer.
Numerals appear only in dates, IDs and timestamps. Indicators read — until there is something true to show.
Technology
- On-premise language model, sized to the hospital’s hardware
- Paragraph-level hybrid retrieval (vector + keyword) with version filtering
- Document ingestion for PDFs and intranet pages, in French and English
- Conflict detection across sources before answering
- Single sign-on and role-based access through the hospital directory
- Query and answer logging for clinical governance
- Evaluation set written with clinicians, rerun on every change
Expected outcome
ExpectedFaster access to the right protocol, the source always one click away, and a system that is trusted because it refuses when it should.
Qualitative and unmeasured. A pilot measures it against acceptance criteria agreed at the start.
What we would measure
- Citation accuracy
- Correct-refusal rate
- Time to the right protocol
- Answers flagged by clinicians
- Coverage of current document versions
Metric names only. Values exist once a pilot measures them.
What a pilot would need
- One department’s protocols, with their version history
- Hardware or a private environment inside the hospital network
- Clinicians to write and review the evaluation questions
- Agreement with clinical governance on logging and review
- Integration with the hospital directory for access control
- A clear statement, agreed with the hospital, of what the system must never do
Promeing ConceptScenario set in Geneva
Building something like this?
Bring the problem in its messiest form. We’ll tell you honestly whether AI is the right way to solve it.