Full-text source
Total records 0With PDF 0Without PDF 0
Total records 0With PDF 0Without PDF 0
Europe PMC and PubMed Central, OpenAlex, Semantic Scholar, Crossref publisher links, Unpaywall, arXiv, and saved full-text links. The local assistant queues missing files automatically and keeps working after this page closes. Original PDFs are preferred; readable XML is retained as a fallback. Identifiers and file contents are checked. Temporary failures retry with backoff; Download retries immediately. Rate limits, paywalls and scanned PDFs may prevent automatic retrieval. DOI, PMID or title lookups are sent directly to these services. The local assistant uses the network of the computer running port 8765, public repositories and available Zotero attachments. If it is unavailable, retrieval falls back to this browser and its network. No AI model is used for downloading. Browser cross-origin rules and website access permissions still apply.
Unpaywall requires a real contact email. Leave blank to skip it. This value is sent only with retrieval requests and is not saved.
Confirm the criterion and correction for each disagreement. All corrections are saved; relevant rules and examples are selected for each subsequent request. The screening model uses this feedback; model weights are not changed. Approval freezes the feedback for formal screening.
Full text comes from Europe PMC (open-access JATS XML) or an open-access PDF (read with pdf.js), is cut into sentence windows (four sentences, step two), and each criterion's most relevant passages are found by contrastive semantic highlighting: cosine similarity to the inclusion rule minus similarity to the exclusion rule, with embeddinggemma (keyword scoring when the embedding model is unavailable). The AI judges each numbered criterion from those passages and the abstract (criteria before and after the passages), quoting a passage. An exclusion needs a criterion not met with a quotation found in the full text; the first criterion not met in the reason order is the primary reason. Two readings per report; direct criterion conflicts and unclear reports go to the reviewer; exclusion review is optional. The fuzzy controller of Rahgozar and Mortezaagha is not used: in their evaluation an all-criteria rule failed half of the true includes.
The approved AI model reads every retrieved full text twice, through selected evidence passages, with two model readings (not independent reviewers). Missing evidence stays Unclear; open the full text to resolve it. One explicitly unmet criterion with a verified full-text quote is sufficient for exclusion, unless the other reading directly contradicts that criterion. Unclear reports and conflicts need human review.