How it works

Verification-first architecture. Five steps, every query. Every claim traceable to a primary source you can open and check.

1. You ask

Natural-language question. Examples: "What is the recommended follow-up for a 6 mm Bosniak IIF renal cyst?" or "Differentiate enchondroma from low-grade chondrosarcoma on MRI."

2. We retrieve

Semantic retrieval across a curated corpus of documents as the primary source. Sources include peer-reviewed radiology and clinical journals, society guidelines (RCR, ESR, ACR, BSIR, BSGAR, and others), openly licensed reference works, and educational atlases. We do not include paywalled content for which we lack a licence.

Where the curated corpus does not contain a directly relevant source, verified web fallback via Exa and Tavily surfaces additional reference material. Web results flow through the same evidence-tier evaluation as corpus retrieval; they are not surfaced unhedged.

3. We rank

Retrieved passages are ranked by both semantic relevance and evidence tier. Tier 1 is guidelines, systematic reviews, and randomised trials. Tier 3 are case reports. Higher-tier sources are preferred where they exist.

4. We synthesise with citations

A large language model (Anthropic Claude family) composes an answer using the retrieved passages, with native inline citations. Every factual claim is anchored to a specific source passage. Synthesis language is calibrated to the evidence: Tier 1 evidence is stated assertively; lower-tier evidence is hedged appropriately.

5. You verify

Every citation is clickable. Open the primary source, check the context, and apply your clinical judgement. MyRadAssistant is a reference tool, not a substitute for primary sources or clinical reasoning.

What we do not do

We do not interpret images. We do not process patient-identifiable information. We do not give patient-specific clinical recommendations and are not for use at the point of care on an individual patient. We are not a substitute for clinical judgement.

Where it runs

UK and EEA-based hosting. Frontend served via UK CDN. Backend on Hetzner (Germany). Vector store on Qdrant Cloud (Germany). Model API calls to Anthropic, OpenAI, Google, and Cohere are governed by Standard Contractual Clauses and the EU–US Data Privacy Framework where relevant. Web fallback queries to Exa and Tavily (US) are similarly covered under the Data Privacy Framework. See Privacy policy for full detail.