About Foreground

How it works

How comprehensive it is

Two streams feed the library, both through the National Library of Medicine's index: the flagship journals read cover to cover (NEJM, JAMA, The Lancet, BMJ, the Annals, and the major surgical journals), and per-specialty surveillance queries refined through actual clinical use. Beyond the curated feed, the search box reaches all of PubMed with full query syntax, or switch it to semantic search, which finds articles by meaning across the library, so you don't need the exact keywords.

Corpus, the research librarian

Corpus is a grounded assistant that lives on every article and across the library. It answers only from the indexed abstracts, plus the open-access full text where available, and cites the papers it used, so every claim traces back to a real article in the library. It reports what the literature says, it does not give medical advice or clinical recommendations. It's a librarian, not a doctor. There are four ways to use it.

• Ask any paper. Question an article's abstract, pull its methods and primary outcome, or draft a PICO critical appraisal, grounded in that paper's own abstract.
• Ask the library. Pose a clinical question and get a synthesized answer that links the real indexed papers it drew from, so you can go straight to the source.
• Journal-club prep. Get discussion questions and the limitations worth probing, grounded in the paper's abstract and your club's own appraisal.
• Explain the statistics. Get a plain-language read of a paper's numbers: effect size, confidence intervals, significance, and NNT, grounded in that paper's abstract.

Corpus is a club feature: sign in (or start a club, free) to ask a paper, ask the library, or prep a journal club. It's included at no charge, with a daily allowance per club to keep it sustainable.

What it doesn't do

No medical-literature tool covers everything. These are the limits worth keeping in mind.

• Relevance is not a quality grade. Behind it, embeddings and a reranker surface the articles most related to a field, sub-topic, or your own question, it measures how central a paper is to the topic, not whether the study is sound. A high relevance score can still sit on a weak trial, so the appraisal is yours.
• The feed favors recent clinical literature. Daily surveillance covers these 32 specialties and their journals, so basic science, niche topics, non-English work, and very new articles are less complete in the feed. The search box covers all of PubMed when you need the rest.
• Summaries come straight from the abstract. The card shows the paper's stated conclusion when the abstract provides one. Nothing is rewritten or invented, but nothing is interpreted either.
• Citation data takes time to build. Counts and impact metrics develop over months, so the newest articles often have little or none.
• It does not replace reading the paper or clinical judgment. It helps you find and organize the literature. The interpretation is up to you.

What it costs

It's free. Start a club, pick your fields, save and organize articles into folders, appraise them, run daily surveillance, and search all of PubMed, with no account needed just to browse. The search bar takes plain language or native PubMed syntax, and tracked topics follow any plain-English question daily, all free, no AI step needed.

How it technically works

For the curious, what's under the hood, and where the computing cost actually lands.

• Every article becomes a vector. An embedding model reads each paper's title and abstract and turns it into a list of ~500 numbers, a point in a high-dimensional space where meaning is geometry. Papers about the same thing land near each other, even when they share no words in common.
• Semantic search is distance. When you search by meaning, your phrase is embedded the same way and the library is ranked by how close each article's vector sits to yours, the cosine of the angle between the two. That's why heart attack surfaces papers that only ever say myocardial infarction: the math knows they point the same direction.
• A reranker does the fine-grained scoring. Vectors are fast but coarse. For specialty tagging and the top of search results, a second model, a reranker, reads the query and a candidate article together and scores how well they truly match. That's what assigns each article its best-fit specialty and lifts the most relevant results to the top.
• Corpus is retrieval-augmented. When you ask a question, it pulls the most relevant abstracts from the indexed library, the same embeddings and reranker above, and answers over them with a fast model. Nothing medical is invented: every statement is traceable to a real paper, and the papers it used are cited back to you.
• The corpus was encoded once, then just maintained. Tens of thousands of back-articles were embedded and reranked up front, that was the one real expense. Each day's new fetch from PubMed is only a few hundred articles, embedded and scored as they land, so the ongoing cost is negligible.
• Most of the pipeline isn't AI. The daily fetch, keyword and full PubMed search, citation data, saves, folders, and the appraisals you write are ordinary, free infrastructure, the National Library of Medicine's index and a small database. The pieces that do call a model are narrow: Corpus chat, Compare, the AI appraisal drafts, the guideline quick-reference cards, and the Radar desk briefings. Those last two are AI-written text from the underlying abstracts, so confirm them against the linked source. All of these run on the operator's key under hard, app-wide monthly ceilings, so total spend stays bounded no matter how many people use it.

Want the whole machine, end to end: the daily pipeline, exactly how papers are ranked, every evidence signal, and where the AI can be wrong? Read the full methodology.