Asclevor Research

Advancing intelligence for medicine.

Our research focuses on making medical knowledge machine-accessible: semantic retrieval over published clinical cases, embedding-based ranking, and careful, honest evaluation of what the systems can — and cannot — do.

Research areas

What we work on.

  • Clinical case retrieval

    Retrieving clinically similar published cases from natural-language questions.

  • Medical knowledge representation

    How clinical concepts, findings, and evidence relate to each other.

  • Embeddings & ranking

    Embedding models and similarity measures for clinical text.

  • Evaluation

    Measuring whether ranked results actually match clinical intent.

Technical work

Notes from the lab.

  • Retrieval

    Measuring relevance in clinical case search

    What does it mean for a retrieved case to be clinically similar? We define the criteria results are ranked against — and document how we measure them.

  • Embeddings

    Embedding published cases at scale

    More than 167,000 patient cases embedded with bge-small-en-v1.5. Notes on throughput, index growth, and what we look for in a model.

  • Evaluation

    Honest evaluation for medical AI

    Retrieval systems fail quietly. We document known limitations alongside measured strengths, and label everything still in progress.

Work in progress — notes are published as the work matures, with methods and limitations included.

Open at the core.

The foundation of Asclevor is a processed dataset of published patient cases drawn from the PubMed case literature. The database is open source — the API is the product.

Cases processed
167,000+
Cases indexed
~40,000
Source literature
PubMed case reports
Source identifiers
PMIDs retained

Responsible AI

Boundaries we hold.

Clinical boundaries
Asclevor supports knowledge discovery. It does not diagnose, recommend treatment, or replace clinical judgment.
Traceability
Every result points back to its published source through its PMID — no anonymous answers.
No patient data
The system is built on published literature. Patient records never enter the pipeline.
Honest evaluation
We report retrieval quality as it is measured today, and label what is still in progress.

Research

Follow the work.

Questions about the research, the dataset, or collaboration — we read everything.