Polaris Automates Multi-Database Literature Searches for More Reproducible Reviews

Every year, researchers around the world spend countless hours copying search strings into bibliographic databases, waiting for results to load, manually exporting records, and then repeating the entire process on another platform. It is one of the least glamorous tasks in science, and yet it underpins virtually every systematic review, meta-analysis, and state-of-the-art synthesis that guides clinical practice, policy, and future research. A new open-source software tool called Polaris, described in the journal SoftwareX, now promises to transform that laborious ritual into a standardized, traceable, and largely automated workflow. Developed by Leonardo Brandão do Prado, Jennifer Vanos, and Jose Benito Rosales-Chávez, the system retrieves bibliographic records from PubMed, Scopus, ScienceDirect, and Semantic Scholar through a single graphical interface, then packages the results in a format ready for reference managers and the screening stages that follow.

The motivation behind Polaris reflects a well-documented problem in evidence synthesis. Literature reviews are essential across disciplines: in the health sciences they track rapidly evolving methods of human care, in geography they have shaped frameworks from economic geography to the geography of poverty, and in interdisciplinary fields such as climate science they assess impacts on ecosystems and social inequalities. There are at least fourteen recognized review typologies, ranging from rapid and scoping reviews to full systematic reviews with meta-analysis. Yet conducting any of them well is time-consuming, particularly during bibliographic database searches, supplementary searches, and the review of each retrieved manuscript for inclusion or exclusion. Existing software such as Covidence and Rayyan helps with screening and selection, while tools like ASySD, Deduklick, BibexPy, and the Systematic Review Accelerator handle deduplication and bibliographic processing. Polaris deliberately targets a different stage: retrieval itself, the point where bias and errors most easily creep in.

Bias in literature searching can arise from the researcher’s own decisions, including selective choices of databases, filters, and search terms. When the systematic process behind a review is not fully disclosed, reproducibility suffers. Errors, meanwhile, can stem from technical and procedural limitations during data retrieval: incomplete metadata, missing author information, export restrictions, or, in the case of Google Scholar, the requirement to visit each paper’s website individually to download metadata. To combat these problems, the developers grounded Polaris in established methodological guidance. The tool follows the eight-stage model of literature searching described by Cooper and colleagues, and it aligns with the PRISMA framework, the widely adopted set of twenty-seven recommendations for reporting systematic reviews and meta-analyses. By standardizing retrieval, automating exports, preserving query logs, and improving the auditability of searches, Polaris aims to reduce potential bias and errors before screening even begins.

Technically, Polaris 1.0 is implemented in R, requiring version 4.5.1 or higher, and built on packages including httr, lite, dplyr, stringr, purrr, xml2, zip, and httr2. The architecture has two layers. The front end is a graphical user interface built with the R Shiny package, where users configure search parameters and monitor progress in real time. The back end is a processing engine with an additional asynchronous execution layer to support long-running searches, organized into modular components for software environment configuration, directory and file handling, database-specific retrieval functions, and the core engine. Users supply their queries in a simple TXT file, select the desired databases, and the software accesses each provider through its official API, retrieves the results, generates harmonized metadata, and saves each query’s output as a Research Information System (RIS) file compatible with Mendeley, Zotero, and EndNote.

Each database demands its own handling, and Polaris implements a dedicated retrieval workflow for every supported provider. PubMed is accessed through the NCBI E-utilities API in three steps: the esearch endpoint counts matching records and retrieves PubMed identifiers in batches of up to 300, the efetch endpoint collects abstracts in XML format parsed with the xml2 package, and the summary endpoint returns structured metadata including title, authors, journal, year, abstract, and DOI in JSON. No personal API key is needed, and a 0.4-second delay between requests keeps the tool compliant with PubMed policies. Scopus and ScienceDirect both rely on the Elsevier API, which requires an API key and an institutional token. Scopus queries are automatically wrapped in the TITLE-ABS-KEY field while preserving the user’s Boolean logic, and ScienceDirect searches title, abstract, keywords, and full text, with abstracts conveniently stored in the description field of the response. Semantic Scholar, which needs no key, uses a bulk search endpoint that supports Boolean expressions, with Polaris translating operators such as AND, OR, and NOT into the provider’s required syntax while preserving the logical structure of the original strategy.

A distinctive strength of the system is its three-level logging design. The first level feeds a real-time monitoring dashboard, updating users on execution status until each search completes. The second, a master log, records comprehensive metadata across all databases: the date and time each query ran, the source database, the query structure, the total number of papers retrieved, the exported RIS filename, and the search status. The third level maintains database-specific logs with the same structure. Together these logs document exactly which queries were executed, where, when, and with what results, making the entire retrieval stage auditable and reproducible. Polaris also handles Elsevier’s authentication errors gracefully: HTTP 401 and 403 errors terminate the affected request and prompt users to verify their credentials, while HTTP 429 rate-limiting responses trigger a temporary pause before retry, all displayed on the dashboard.

To demonstrate reliability, the team compared Polaris retrieval counts against the official web interfaces of PubMed, Scopus, and ScienceDirect across 60 queries, 20 per database, using identical search strings and equivalent fields. Agreement was assessed with linear regression, Pearson’s correlation, Bland–Altman analysis, and relative percentage differences. The results were striking: a Pearson correlation coefficient of r = 1.000 and a coefficient of determination R² = 1.000, with points closely following the identity line across searches ranging from a handful of records to several thousand. The Bland–Altman mean difference, or bias, was just −7.92 records, a negligible proportion for searches returning more than 5,000 results. Nearly all observations fell within the 95% limits of agreement. The mean relative difference was −0.15% and the median 0.00%, meaning most searches produced identical counts. PubMed and Scopus showed almost complete agreement, while ScienceDirect displayed a small number of larger but still limited deviations. A log-transformed Bland–Altman analysis estimated a bias of only −0.0007 log units, confirming that discrepancies do not grow with search size. The complete validation dataset and R scripts are publicly available on GitHub and archived on Zenodo.

The software does have limitations tied to each provider’s API. Without a key, PubMed caps results at 10,000 per query; ScienceDirect allows 6,000 under an API and institutional key; Scopus imposes no limit on broad retrievals; and Semantic Scholar is limited to roughly the first 300 results. Validation for Semantic Scholar was not possible because its webpage and API return different results, and for queries hitting a provider’s maximum, Polaris was considered to have correctly handled the constraint rather than achieving perfect agreement. Future versions plan to add front-end filters for year, field, and study design, along with support for the IEEE and SciELO databases to broaden coverage to engineering and Latin American literature.

Early applications suggest Polaris can scale to demanding real-world projects. In a critical review of outdoor workers and heat exposure in the United States, the software handled approximately 100,000 results drawn from multiple databases. In a systematic review of human attitudes and behavior toward snakes, queries using Boolean operators, parentheses, and asterisk truncation retrieved 21,734 records in preliminary results, demonstrating robust handling of complex syntax and special characters. Because repetitive tasks such as copying queries across platforms and manually exporting results are automated, and real-time monitoring catches interruptions during long runs, the developers argue that Polaris shifts literature retrieval from a manual, platform-dependent chore toward a standardized, traceable workflow. Its MIT-licensed R and Shiny code base makes it straightforward to adapt, extend, and reuse, potentially changing how evidence synthesis is conducted day to day in an era of ever-expanding publication volume.

Subject of Research: An automated, open-source software system for multi-database bibliographic literature retrieval supporting reproducible systematic reviews

Article Title: Polaris – an automated literature retrieval system

Article References: Polaris – an automated literature retrieval system. (n.d.). https://doi.org/10.1016/j.softx.2026.103024

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103024

Keywords: Polaris, literature retrieval, systematic reviews, bibliographic databases, PRISMA, open-source software, R Shiny, PubMed, Scopus, ScienceDirect, Semantic Scholar, reproducibility