Literature review workflow
Which AI drafts a literature review from my papers with quotes, pages and BibTeX?
Updated by the Vidi team
Plans start at $5 a month.
You want four things in one pass. The AI should read only your papers and pull out the evidence as quotes with page numbers. It should draft a review from that evidence. Then it should hand you something that compiles in Overleaf with working citations.
You can get this from one tool or from a chain of tools. Either way, the slow part is the same. You still need to check that each quote is really on the page it claims. The draft also needs your own argument.
This guide covers what a useful evidence table looks like and how to do the job with Vidi. It also shows how to do it without Vidi, using Elicit and a reference manager. Finally, it covers how far AI extraction can be trusted.
How it works
How to draft a literature review with Vidi Research
Build your library
Sign in to Vidi Research and add your papers. You can upload PDFs or add papers by DOI, arXiv ID, OpenAlex or PubMed ID, a link or a title. If you already keep references elsewhere, import them as a BibTeX or RIS file. That way your library starts from the records you trust.
Choose papers and a question
Pick 2 to 12 papers and write a research question. Vidi builds an evidence table with quotes and page numbers. It maps agreements, disagreements, gaps and trends, then drafts a literature review section. This takes about 10 to 20 minutes and costs 5 credits. Reviews are written in English.
Check, then export to Overleaf
Open each source PDF and search for every quote you plan to rely on. Confirm the page and the meaning. Then export the review as CSV, as Markdown, or as an Overleaf project with LaTeX and BibTeX. You can also export your library as references.bib.
What a useful evidence table and draft look like
Give each claim its own row, not each paper. A good row has the claim in your words and the verbatim quote that supports it. It also has the page number and the paper's citation key. Add a column that says whether the quote supports the claim directly or only partly.
Decide which page numbering you use: the journal's printed page numbers or the PDF file's page count. They can differ, so record your choice in the table header.
For the draft, organise by theme or by question, not one paragraph per paper. A paper-by-paper summary is an annotated bibliography, not a review. Use the agreements, disagreements and gaps in your table to structure the sections. Then add what a tool cannot supply for you: why the disagreements exist, which studies you weigh more and why, and what your own work adds. Treat any AI draft as scaffolding for that argument.
Doing it with a chain of tools instead
If you prefer not to use an all-in-one tool, split the job in four. First, use an extraction tool for the evidence table. Elicit's Systematic Review workflow lets you upload PDFs or select papers already in your library. It fills a table with data from each paper along with quotes and reasoning. You can define the columns yourself or let Elicit suggest them. Its stages are Gather, Extraction and Report, and the Report stage generates a research report. The sources used here do not confirm whether Elicit's quotes include page numbers or whether it exports BibTeX. Check its current options before you commit.
Second, keep one reference manager as the single source of your .bib file and citation keys. Third, check page numbers by hand in a PDF reader. Fourth, write the review in LaTeX using those keys. Then put the .tex and .bib files into your Overleaf project. This takes longer, but every citation comes from a record you control.
How far you can trust AI extraction
AI extraction is useful but not finished work. A feasibility study of Elicit for systematic-review data extraction compared two runs from different user accounts. It found 90% agreement on extracted values. Some supporting quotes were missing in one run, the other or both. The same study reports that Elicit helped find a small number of errors in the human-made reference data. So AI can catch mistakes as well as make them.
In practice, pilot on a small set before running everything. Elicit itself runs a pilot extraction on 10 included papers chosen at random, so you can check values and quotes first. Copy each quote and search for it in the PDF. If it does not turn up word for word, treat it as a paraphrase. Do not let an AI write BibTeX entries or DOIs from memory. Export them from your reference manager or library instead. Finally, a correct quote only shows where a claim comes from. It does not prove the claim reads the paper correctly. That judgement stays with you.
When another tool is the better choice
Vidi Research works on 2 to 12 papers per review and writes reviews in English. That suits a related-work section, a thesis chapter draft or a grant background built from papers you already have.
For a formal systematic review, Elicit's Systematic Review workflow is the better fit. It covers search, title and abstract screening, and full-text data extraction. Elicit is also clear about its own limits. For publication or regulatory-grade systematic reviews, it says you will often have to default to traditional keyword-based search methods. Whatever tool you use, run and document a proper keyword search alongside it if your review must be reproducible.
If you must write the review in another language, Vidi Research is not the right tool for the drafting step. If you already have a working reference manager and a writing habit you like, the chain of tools above may suit you better than switching.
Vidi Research and Elicit for a review from your own papers
| Vidi Research | Elicit (Systematic Review workflow) | |
|---|---|---|
| Works from your own papers | Yes. You choose 2 to 12 papers from your library. | Yes. Upload PDFs or select papers already in your library. |
| Evidence table | Quotes and page numbers for each piece of evidence. | Extracted values with context (quotes and reasoning). Page numbers not confirmed in the sources used here. |
| Draft or report | Maps agreements, disagreements, gaps and trends, and drafts a literature review section. | Report stage generates a research report. |
| Export | CSV, Markdown, or an Overleaf project with LaTeX and BibTeX. Library exports as references.bib. | Not confirmed in the sources used here. Check current options. |
| Formal systematic reviews | Not built for screening. Limited to 12 papers per review. | Covers search, title and abstract screening, and full-text extraction. Pilots extraction on 10 random papers first. |
| Search | Questions search OpenAlex and arXiv and return the three closest papers. | Includes search. Elicit says publication-grade reviews often need traditional keyword search too. |
| Language and cost | Reviews in English. 5 credits per review. | Not covered in the sources used here. |
Limits worth knowing
- AI drafts can be wrong. A quote shows where a claim comes from, not that the claim reads the paper correctly. Check every quote, page and interpretation against the PDF before you cite it.
- Vidi Research reviews cover 2 to 12 papers and are written in English. It is not a systematic-review tool with screening, so a formal review still needs a documented keyword search.
- Page numbers can mean the printed journal page or the PDF file page. Confirm which one your table uses before the citations go into your text.
- An AI draft is not a finished review. The critical synthesis, the weighting of evidence and your own contribution have to come from you. Check your university's or journal's rules on disclosing AI use.
Sources
Can I make the AI use only my own PDFs?
Yes, with tools built for it. In Vidi Research, the review is built from 2 to 12 papers you choose from your library. In Elicit's Systematic Review workflow, you can upload PDFs or select papers already in your library. With any tool, check where each claim in the output comes from.
Will the AI make up quotes or page numbers?
It can get them wrong. A feasibility study of Elicit found some supporting quotes missing across two runs. Copy each quote and search for it in the PDF. Confirm the page and read the surrounding text to check the meaning. Do this before the quote goes into your review.
How do my citations end up working in Overleaf?
Vidi Research exports an Overleaf project with LaTeX and BibTeX. If you build the review with other tools, keep one .bib file from your reference manager and cite its keys in your .tex file. Then add both files to your Overleaf project. Avoid BibTeX entries written by an AI from memory.
Can I bring references I already have?
In Vidi Research, yes. You can import BibTeX or RIS files and export your library as references.bib. If you use another workflow, export from your reference manager. That keeps one source for your citation keys and helps you avoid duplicates or mismatched keys.
How much does a literature review cost on Vidi?
A literature review in Vidi Research costs 5 credits, and one credit is worth one US dollar. New accounts get 5 free credits, with no credit card needed. Asking research questions is free, up to 20 a day.
Is this good enough for a systematic review?
Not on its own. A systematic review needs a documented search and screening process. Elicit's Systematic Review workflow covers search, screening and extraction. Even so, Elicit says publication-grade reviews often need traditional keyword search as well. Use AI to speed up extraction, not to replace the method.
Should I upload unpublished manuscripts?
Read the tool's data and privacy policy first, whichever tool you use. This page does not cover how any tool stores or uses uploaded files. If a manuscript is confidential or under review, check with your co-authors or the journal before uploading it anywhere.