Literature Review

Open Literature Review and type one research question, for example Relationship between the gut microbiome and Parkinson disease. Choose the review’s thinking level in the question box: Low, Medium, High or Max, whatever your chat uses. Unless you pick one, the review starts at its model’s default, which is Medium today. Every step of the review runs at the level it started with.

• Low: the least thinking at each step.
• Medium (default): about 11–16 minutes and 1,100–1,200 credits.
• High: more care on hard steps; its time and credits fall between Medium and Max.
• Max: the deepest thinking, about 30–34 minutes and 2,400–3,000 credits.

The run moves through these stages. The current one is shown as it works:

1. Planning the search: the question becomes inclusion and exclusion criteria and a few sub-questions, each made of concepts with their synonyms and MeSH headings.
2. Searching the literature databases: the RBAi corpus of 11 million papers and OpenAlex by meaning, and PubMed, Europe PMC and OpenAlex with Boolean queries built from the concepts. MeSH headings are checked with the NLM, and each PubMed query is checked by its number of hits before it runs.
3. Merging duplicates and finding abstracts: a paper that came without an abstract is looked up in PubMed, Europe PMC, Crossref and Semantic Scholar.
4. Adding the citation network: papers citing and cited by the best matches join the candidates.
5. Fetching full text: open full text with its licence, from PMC and Europe PMC among others.
6. Screening each paper against the criteria: the model includes or excludes each candidate on its title and abstract, with a reason. A paper with no text can only be listed as related.
7. Reading the included papers: the model reads each one, in full text when it is open, and writes notes in its own words, with short quotes from the paper. A review reads up to 60 papers; when more meet the criteria, the model first chooses which to read, from its screening notes and the question, and says why.
8. Writing the review: from the notes and abstracts, going back to a paper’s text when a note is not enough. The writer adds diagrams and tables where they help a reader, such as a mechanism, a comparison of studies or how the studies were screened, drawn with the review’s own counts.

Citations come from each paper’s reading notes and are not checked separately: since 30 October 2026 a review has no pass that reads each cited statement against its papers after the review is written. Reviews checked before then still show their check’s counts.

When figure reading is on and the review’s model reads images, RBAi also reads the figure images of papers whose PubMed Central copy allows it, while the papers are read. Figure readings

Show steps opens each stage’s reasoning, and Stop ends the run and keeps what it has written so far. Reviews that are still running are listed on the Literature Review page and in Jobs.

A long review is not cut at a time limit: before each 15 minutes run out, it saves its place and continues in a new run, and the steps show Continuing in a new run. When your credits run out, the review stops, keeps what it found and wrote, and offers more credits.

The review opens with its title and a line giving how many papers were read (in full text and from the abstract) and how many are cited; a review checked before 30 October 2026 also gives how many cited statements were confirmed by their papers. Reviews with figure readings add a line such as Figures read: 30 of 62. Only papers the review read can be cited: a paper with no abstract or full text is listed under Not read as evidence. Findings cite numbered papers, like [1, 4]. Hover over a number, or tap it on a phone, to see the paper:

• its title, first author, journal and year;
• a Full text badge when RBAi found the paper’s full text;
• PMC, DOI and PubMed links;
• Ask about this paper.

The text can hold diagrams and tables that the writer chose. A diagram’s node or arrow and a table cell that state a finding cite their papers like a sentence. A diagram that cannot be drawn shows a short note and its source.

After the text come the references, which list only the papers the text cites. On a computer, clicking a number jumps to its reference. Each reference has Reading notes: the reviewer’s notes on the paper, each with the paper’s own words behind it. Reviews from before 28 October 2026 show Extracted data instead. Last, collapsed, Not read as evidence lists the related papers that had no text and the web sources that are not scholarly articles. A web result that is an article, because its address or its page gives a DOI, PubMed, PMC or arXiv id, is screened and read like the papers the databases found.

Each review keeps the run’s own counts: records found and duplicates merged, candidates screened, included and excluded (with the reasons), and papers read in full text and from the abstract. When a flow diagram helps, the writer draws one from exactly these counts.

Search details in the ⋯ menu lists the records from each source, the unique papers each source added and the run’s steps.

The model reads each included paper, its full text when it is open and otherwise its abstract, and writes notes in its own words: the aim, the system and sample size, the design and methods with their conditions, the exposure and comparator, the outcomes, the results with their numbers, the main finding, the authors’ conclusion, the limitations, the quality of the evidence and, for a question about an effect, the paper’s stance. Each note carries short quotes from the paper, checked to occur in its text; line breaks, hyphenation, ligatures and typography are ignored in that check.

Download evidence table in the ⋯ menu saves one row per paper read as CSV: design, population or system, sample size, intervention or exposure, comparator, outcomes and main finding, from the notes, with each cell’s quote and where it is in the paper. References (BibTeX) and References (RIS) save the references for a reference manager.

Reviews from before 28 October 2026 have a papers table in 11 fields instead: study design, sample size, population or model, intervention or exposure, dose, comparator, primary outcome, effect size, statistics, key findings and limitations.

In those reviews each value comes with the paper’s own words that state it, in quotes. The server kept a value only when each of its quotes occurs in the paper’s text and every number in the value appears in its quotes.

What is read: for a full text, the passages its search ranks first for each of the 11 fields, up to 12,000 characters a paper; a full text that fits in that is read whole. For a paper with only an abstract, the whole abstract. Each row says what was read.

Evidence level and relevance are the model’s assessments, not values from the paper, so they have no quotes. Figure readings are not used to fill the table: each row only says which of the paper’s figures were read.

In the CSV export, each field’s column is followed by a <field>: quote column with that value’s quotes, separated by |. The Extracted from column says what was read for the paper, and the last column lists the figures read from their images, each with its licence and PMC link.

When figure reading is on, a review reads the figure images of its full-text papers, not only their captions, if the paper has a copy in PubMed Central whose licence allows commercial use: CC0, CC BY, CC BY-SA or CC BY-ND. The images come from the PMC Cloud Service of the US National Library of Medicine. Figures under other licences, such as CC BY-NC, are counted and skipped.

An AI vision model reads each image with its caption: the panels, axes and units, groups and sample sizes, and the direction and size of each effect. Values read off a bar, a point or a curve are approximate. Each reading is kept with the review as a passage headed AI reading of the image of Figure N (approximate; check the figure), with the figure’s PMC page, its licence and its copyright line. No copy of the image is kept.

• The line under the title gives the coverage, such as Figures read: 30 of 62 (32 skipped: licence does not allow commercial use), or (time limit) when the review’s time ran short.
• In the chat, an answer that uses a reading says so, for example “from an AI reading of Figure 2 of [3]”, and links the figure’s page on PMC.
• In reviews from before 28 October 2026, each reference’s Extracted data lists the figures read, with their links, licences and copyright, and how many were not read and why.
• Readings are not used in the papers table.

RBAi can switch figure reading off. A review made while it is off reads the captions only and shows no Figures read line. A reading that fails never fails the review; it is counted in the coverage line. The readings use the review’s credits.

Open the review’s chat with Ask to ask about its papers. You can also start from Ask about this paper on any reference. A question is charged by what its model calls cost.

The chat answers from everything the review read. Each paper is kept as passages that follow the article (its sections, each table and each figure caption), the whole article up to 400,000 characters, or its abstract when that is all there is. A question is matched to the passages by meaning and by exact words together, then a reranker orders the best 50, and the top 25 passages go with the question, with a one-line list of the review’s papers. When those are not enough, the chat reads more itself: a paper’s reading notes, another search of the passages, or more of a paper’s text. A phrase in double quotes must appear word for word.

Name papers in the question, such as [3], [3, 7], [3-5] or an author and year, or use Ask about this paper, and the question gets those papers’ reading notes and their 25 passages most relevant to it. Every paper states the text the chat has of it: full text, full text cut at a number of characters, abstract only, or no text. So a missing point reads “not in the text available for [n]”, with the reason, never “the paper does not report it”. When you ask about a paper that has no text yet, the chat looks it up once in PubMed, Europe PMC, Crossref, Semantic Scholar and OpenAlex and keeps what it finds: the abstract, or the open-access full text where its licence allows. If none exists, the answer says so and gives the paper’s links.

Every [n] in an answer is checked against the papers the chat was given. A number that is not one of them is removed, and a line under the answer says how many were removed. Right after a review finishes, a line may say that some passages are still being indexed: until then they are found by their words only.

Visual summary (Beta), beside Ask in a finished review, draws one picture of the review’s story in a journal, slide, poster or social format, planned and reviewed by AI. It opens the same viewer as in Studio, with the review’s newest visual summary and its Edit with AI, or the formats to draw one. Visual summary

The ⋯ menu has these once the review has finished:

• Save as PDF…
• Download Word
• Download Markdown
• Download evidence table: the evidence table as CSV, with each cell’s quote (Reading notes and the evidence table)
• References (BibTeX) and References (RIS): the cited papers, for Zotero, EndNote or Mendeley
• Copy text
• Open in Report Writer
• Edit and publish: the review becomes a Report Writer page, with its citations and references, for you to edit and then publish (Website)
• Send to an RBAi member: a Connect message with the review’s title and a link only the two of you can open (Send work to a member)
• Search details (PRISMA flow)