How it works and limitations
quedicelaciencia summarizes what the scientific literature says about a question. It searches for papers, automatically classifies what each one answers based on its abstract, and weighs the evidence by study quality.
1. From your question to a search
- If the question is in Spanish, it is translated into English: first from your browser with MyMemory; if that fails, on our server with Cloudflare Workers AI.
- The concepts are extracted and synonyms from MeSH (the US National Library of Medicine thesaurus) and Wikidata are added. A model filters them so that only true synonyms remain.
2. Searching for papers
- Sources: Europe PMC (biomedicine, includes PubMed), OpenAlex (all disciplines), Semantic Scholar and, for physics, mathematics and computer science, arXiv.
- Only papers with an abstract are used. If the same paper appears in several sources, the records are merged.
- The quick search analyses about 100 papers, starting with the strongest designs (systematic reviews, trials) and the most cited. The exhaustive search continues with the rest, the references of systematic reviews and related topics.
3. Classifying each abstract
Each abstract is classified by Jev, a decision model by TypeSafe (not a chatbot; it does not generate text). It decides:
- whether the paper reports its own findings that answer the question;
- what it answers: yes, no, partial, no effect or inconclusive;
- the study design (meta-analysis, randomized trial, cohort…);
- whether it studies humans.
For physical-science questions, adapted questions about measurements, experimental limits and simulations are used.
4. Verdict
- Each paper is weighted by its design (a systematic review weighs more than an observational study), its relevance and the confidence of the classification. In health questions, animal or in vitro studies weigh less.
- Reliability corrections:
- retracted papers do not count;
- many papers from the same group of authors count as fewer;
- if studies are divided and there are many systematic reviews, the headline follows the reviews.
- The headline reflects which side gathers more weight; the certainty reflects how much evidence there is and how consistent it is.
- The plain-language summary is written from templates (not by a generative model) and cites the studies it relies on.
What it is NOT
- It is not a systematic review. It is an automatic summary, useful to get oriented, and it can be wrong.
- It only reads abstracts, not full texts, and abstracts sometimes exaggerate or oversimplify.
- It is not medical or professional advice. For health decisions, consult a healthcare professional.
Known limitations
- Publication bias: positive results are published more often than negative ones.
- Language: almost all indexed literature is in English.
- Coverage: many publishers do not share their abstracts, and the sources have usage limits. Some relevant papers may be missing.
- Automatic classification: Jev makes mistakes, especially with ambiguous abstracts. If you spot one, use “Wrong classification?”: it helps us improve.
- The question matters: a vague question or ambiguous terms give worse results.
Credits
Thanks to Europe PMC (EMBL-EBI), OpenAlex (OurResearch; its data include Unpaywall's), Semantic Scholar (Allen Institute for AI), NCBI (MeSH and E-utilities) and Wikidata. Thank you to arXiv for use of its open access interoperability. Classification by Jev (TypeSafe). Hosted on Cloudflare. Open source under the MIT license.