TL;DR: Today, I’m releasing ClinicalEncoder25, an interpretable non-generative reasoning model that understands clinical texts at millisecond speed, with token-level precision. Built on the new Diagnosable ColBERT architecture, it maps every word to a semantic clinical graph, enabling real-time reasoning and retrieval. It also enables debugging, with every token becoming an opportunity to uncover misunderstandings and potential mistakes, providing a level of interpertability never seen yet in ColBERT models. Try the live demo, explore the model on HuggingFace, and join me at the AI Thinkerers meetup in Brussels [December17] for a deep dive!
Stop generating, start understanding!
Most AI labs today are fixated on generation. But before generating, you must understand; and most AI models today only scratch the surface when it comes to true clinical understanding. No more!
Since my PhD, I’ve pursued a vision: an AI map of healthcare, a digital atlas that grounds AI reasoning in structured medical knowledge, generalizing across ontologies, scientific literature, and all forms of clinical communication.
After two years of meticulous work, I’m thrilled to unveil the first version of this map and, with it, ClinicalEncoder25: an encoder that ingests clinical documents in milliseconds, extracting insights without secondary models for medical entity recognition or linking. Whether you need highly expressive vector embeddings or ontology-grounded reasoning, ClinicalEncoder25 will deliver for your business!
The Diagnosable ColBERT: Interpretable by Design
Late-Interaction Retrieval, Reimagined
ClinicalEncoder25 isn’t just another encoder. It performs late-interaction retrieval, clinical coding (UMLS, SnomedCT, or any other ontology), and topic extraction—all from the same representations. Unlike every other clinical encoder released before, this models knows that “PAPA Syndrome” is an X-linked interleukin-related deficiency, and can retrieve relevant PAPA documents even with generic queries like “interleukin deficiency,” no complex augmentation needed.
Hallucination-Free, Token-Level Reasoning
Unlike LLMs, ClinicalEncoder25 represents entire documents in a single pass, in milliseconds, without generating tokens or hallucinating. It connects the dots: if a patient works at a car repair shop and is later noted to have lead contamination, the model infers “past history of exposure to lead-based paint,” augmenting its representation with all available evidence.
Unlike static tagging models, it can also accurately retreive concepts for which no individual entry exists in a predefined ontology, combining any number of pre-existing concepts freely to produce new, adhoc vector representations for any information the model reads (this can slightly lower interpretability, but means that retrieval doesn't have to suffer from limited ontologies).
Diagnosable by Design
Traditional ColBERT models are interpretable only in hindsight. ClinicalEncoder25 changes that with its Diagnosable ColBERT architecture: every token is directly interpretable, mapped to a semantic clinical graph.
You can verify what the model understands immediately, without search queries, both at the mention level (“ranitidine”) and at the global semantic level (“no known allergy to ranitidine”).
Try it yourself: Live Demo Hover over any word to see real-time relationships and concept mappings.
The ClinicalMap 2025: A New Atlas for AI in Healthcare
From BioLORD to ClinicalMap
In 2022-2023, my BioLORD models set a new standard for biomedical embeddings, with nearly 100,000 monthly downloads. But to achieve true clinical reasoning, we needed more than co-occurrence and synonyms. More than basic contrastive learning. We needed structured, foundational representations.
Breakthroughs in Knowledge Cartography
ClinicalCartographer 2025 combines knowledge graphs, sparse representations, and synthetic data generation, to produce an unprecented AI-ready clinical maps. The result is a vector-labeled map of SnomedCT concepts, designed for use with ClinicalEncoder25.
While the map is released under a CC-BY-NC license (and likely requires an additional SnomedCT license for use), the underlying ClinicalCartographer model is ontology-agnostic: it is capable of mapping to UMLS, ICD10, or even custom clinical ontologies containing brand new concepts. Reach out to us if that's something you would like to try!
What’s Next?
Join Me in Brussels: AI Thinkerers Meetup (17/12/2025)
I’ll be presenting ClinicalEncoder25 and the Diagnosable ColBERT architecture at the inaugural AI Thinkerers meetup in Brussels. We’ll pop the hood on the model, walk through live Jupyter notebooks, and demonstrate how to debug token-level reasoning—all without generating a single token!
Talk Title: Stop generating, start understanding! Introducing the Diagnosable ColBERT architecture for late-interaction retrieval and reasoning on clinical texts
Multilingual models are also planned soon, in Q1 2026; a Dutch version is already pre-trained, and French+German+Spanish are of course coming as well. Get in touch if there are languages that you care about particularly!
Get Involved
- Try the demo: http://demo25.parallia.eu/
- Explore the model: HuggingFace
- Attend the talk: AI Thinkerers Meetup Brussels (December 2025)
- Reach out: For custom maps, collaborations, or questions, contact me!