Sources and search

The backbone sources

Every search records which sources answered, which failed and which were not searched, so a systematic review can report its strategy honestly. Status chips next to results show this per source.

Searching

1. Sources tab → type a question or a Boolean query.
2. Filter by year, type and open access; sort by relevance, date or citations.
3. Tick results and Add to project; duplicates across sources are merged into one work.

Systematic search plans

Plan turns a question (and the review PICO, if set) into concept blocks with synonyms and one query per database: Boolean for OpenAlex, PubMed and Europe PMC, keyword for the others. Edit any line, run everything, and export the strategy for the methods section.

Snowballing

From the works already in the project, find what they cite and what cites them (OpenAlex). Candidates are ranked by how many of your works connect to them; add the ones that fit.

Full text

Open-access copies are fetched automatically and stored with their provenance and rights class. Structured XML comes first — the JATS full text that PubMed Central / Europe PMC, Europe PMC preprints, bioRxiv/medRxiv, PLOS, Frontiers, PeerJ, MDPI and Copernicus publish — because it parses into clean sections, tables and references; PDFs (Unpaywall and OpenAlex locations, Europe PMC, arXiv) are the fallback and are converted to TEI XML with GROBID, with OCR for scanned files. Works without an open-access copy are listed so you can upload your licensed copy.

The rights class a copy carries decides what may be done with it, and the Files list says so beside each file: an open-access copy is read by your connected AI in full; a copy you upload is read by your AI straight away — your upload is the statement that you are entitled to work on it here, there is no attestation step, and your name is recorded as the uploader; a copy whose rights are not recorded is limited to a short excerpt. Uploads are still not exported and still not handed to another project without a rights decision. Open the file to see every operation's answer. See Your AI agent → What your agent may read.

Requests — what the assistant could not obtain

When Hikma cannot get a full text, it does not ask you in a chat message. It tries every authorised route first — PubMed Central and Europe PMC structured XML, the publisher's XML, then open-access PDFs — and only what is left becomes a request on the project's Requests tab.

Each row carries what you need to find the paper yourself: title, authors, journal, year, volume and pages; the DOI, PubMed and PMC identifiers as links; the publisher or open-access address when one is known; why it is needed and what depends on it (evidence units, claims, bound numbers, sections); and what Hikma tried, route by route, with the time and the outcome. Upload the file from the row and the request closes itself; or answer “I cannot obtain this” with your reason, which your connected agent reads back and works around.

PRISMA

Counts update live: records identified per source, duplicates removed, screened, excluded by reason, assessed in full text, included. The diagram exports as SVG or PNG for the paper.

Risk of bias by hand

Review › Risk of bias lists every study included at full text. Choose the instrument for the review — RoB 2 (randomised trials), ROBINS-I (non-randomised interventions), Newcastle–Ottawa (cohort/case-control), QUADAS-2 (diagnostic accuracy), the JBI checklists (cross-sectional, case series) or AMSTAR 2 (overviews of reviews) — and click a study. Each domain shows its signalling questions where the instrument has them; the answers propose a judgement, but the reviewer chooses the judgement and writes the supporting text (required for domain-based tools). The overall rating follows the instrument's rule and is shown live. Save and move to the next study.

The traffic-light table and the weighted bar chart update as you go and export as SVG or 300-DPI PNG; “Insert table” adds the risk-of-bias table to a manuscript as tracked changes. For RoB 2 and ROBINS-I, a connected AI agent can fill a first pass from the extracted evidence (and, where an administrator has switched platform AI on, so can the “Propose with the assistant” button); nothing is stored as final until a reviewer saves it.

GRADE and the Summary of Findings

Review › GRADE rates the certainty of evidence per outcome: start from the design (randomised trials high, observational low), rate the five domains (risk of bias, inconsistency, indirectness, imprecision, publication bias) as not serious, serious or very serious, add upgrades for observational evidence, and read the certainty with its derivation. Enter the Summary-of-Findings fields (studies, participants, effect — typed or taken from a meta-analysis run — absolute effects, importance) and the plain-language statement; the table exports and inserts into the manuscript.