Grounded AI search over PubMed and the biomedical literature

Most tools that put AI on top of PubMed are good at summarizing and careless about sources. For real research that is the wrong trade. medground grounds answers across the biomedical literature, from PubMed and curated databases, so every claim traces back to a paper you can open. It is not limited to one field.

It grounds the whole literature, not one topic

medground builds a corpus you control and answers only from what is in it, with every citation verified. The corpus is not fixed to a single specialty. You can ingest broad coverage from PubMed across any area of biomedical research, and layer in curated, structured sources on top:

• PubMed for broad, current literature across fields, from immunology to neurology to cardiology and beyond.
• CIViC for curated, leveled biomarker evidence, as one structured layer among others.
• More sources over time, so the same grounded workflow widens as the corpus grows.

The oncology examples elsewhere on this site are just that, examples. The method is general: whatever you ingest, medground retrieves from it and checks every citation.

Why "AI on PubMed" usually disappoints

A general assistant asked about PubMed will write a fluent answer and attach references that may or may not exist. It can cite the wrong study, paraphrase past what a paper claimed, or invent a plausible DOI. The reading is fast; the provenance is not trustworthy. For a literature search where the citation is the deliverable, that is a dealbreaker.

How medground stays honest

The discipline is the same regardless of source or field:

• Retrieve from a finite, inspectable corpus you built, not from model memory.
• Write each claim as a discrete statement carrying a paper ID.
• Run check_grounding, a deterministic gate that rejects any claim whose citation is not real and retrievable.
• Open the cited paper to confirm the reading yourself.

Build a corpus across the sources you need

Pull a curated layer, then widen with fresh PubMed papers on your topic:

# curated structured evidence
uv run medground ingest civic

# broad literature: pull fresh PubMed papers on a topic
uv run medground ingest pubmed --query "your research question"

The corpus grows as you ingest more, so a prior answer can shift as new evidence lands. That is the point: the system reflects what is actually retrievable, not a snapshot of a model's training data.

Frequently asked questions