Keeping Medical Research Tied to PubMed Sources

MedIntel retrieves PubMed records and drafts a research summary from selected abstracts. It is an independent reference system for literature review, not clinical decision support or treatment advice.

Keep the PubMed call off the event loop

The search route validates the query and retrieves records through Biopython's synchronous Entrez client. FastAPI runs that call in a worker thread so waiting for PubMed does not block the event loop. The thread still remains occupied until the retrieval finishes.

The parser turns each result into a PMID, title, abstract, journal, publication date, authors, and PubMed URL. It skips a record it cannot parse and continues through the results. Keeping those fields together gives the reader a specific article to open when checking a summary.

Be precise about what the model reads

Synthesis accepts between one and ten article records and trims each abstract to 1,500 characters before constructing the prompt. The request example below includes the PMID and URL in each source block. The article list comes from the browser; a typed payload alone does not prove that the text was retrieved from PubMed.

This is abstract-based review. It does not retrieve the full methods, tables, or supplementary material from each paper. Trimming can also remove part of an abstract, so a finding in the draft needs checking against the complete source.

The prompt asks for source-by-source appraisal before combining findings. Reported statistics and disagreements belong in that summary when the supplied text contains them. The request format does not establish study quality or make the selection a systematic review.

Separate the gate example from the current build

The gate example below checked a request counter and a stored monthly spend value. It bypassed both checks when Redis credentials were missing. The screenshot shows the lookup, not how spend was recorded or reserved.

October 13, 2026 update: the public build applies request throttling when Redis is configured. It has no monthly spend check. Titles, journals, and abstract excerpts enter the prompt; PMIDs and links stay in the interface's article records. The screenshots are design examples, rather than a snapshot of this build.