How it works Overview

OpenAlex gathers the world’s research output from thousands of sources, organizes it into one connected knowledge graph, and shares it with everyone for free. This page is the map of how that pipeline works, from broad strokes. Each step is documented in depth in the “How we build it” section of the relevant entity page, and this page links into them as it goes.

Gather

When a researcher publishes an article, book, or dataset, information about it is registered with agencies like Crossref and DataCite, or deposited in institutional and national repositories. OpenAlex pulls records from these sources continuously — the database evolves hourly. The catalog was seeded by the discontinued Microsoft Academic Graph (MAG), whose final open dataset OpenAlex adopted in 2021.

We track the external indexes a record can come from — Crossref, PubMed, DataCite, DOAJ, arXiv — and the venues those records name become sources: journals, conference series, ebook platforms, and repositories. The core inputs we pull from today include:

• Crossref and DataCite, the DOI registration agencies
• ORCID and ROR, the open identifier systems for researchers and institutions
• PubMed and PubMed Central
• DOAJ and the ISSN International Centre
• Unpaywall and the Internet Archive
• Aggregators and subject repositories: HAL, arXiv, Zenodo, Dergipark, OSTI, RePEc, and many more
• Thousands of institutional repositories, from UNC’s CDR to Michigan’s Deep Blue (full list)
• Parsing of 60M+ open access PDFs, journal landing pages, and direct publisher feeds
• Corrections from users through community curation

Organize

The records flowing in are messy and redundant, so the heart of OpenAlex is turning them into a clean, connected graph of entities. Some of those entities are native — OpenAlex mints its own IDs by making judgment calls about fuzzy real-world boundaries — and some are a consistent vocabulary wrapped around things that already exist crisply; the Entities overview explains that distinction and how much to trust each kind of ID.

Records become works. Each incoming record is matched (by DOI or other metadata) against the works already in OpenAlex. If it matches, the record enriches the existing work; if nothing matches, it may seed a new one. Duplicate records of the same work are merged into a single work — the core node that every other entity connects to.

Authors get disambiguated. The name strings on a work (“J. Smith,” “John A. Smith”) are clustered into real people, each assigned a stable author ID. This is author disambiguation, a machine-learning process that weighs name, co-authors, affiliations, topics, citations, and ORCID.

Affiliations get matched. The free-text raw affiliation strings on each work (“MIT, Boston, USA”) are parsed and linked to ROR-backed institutions and countries — which in turn feeds the affiliation signal used in author disambiguation.

Works get classified. Text classifiers read each work’s title and abstract to decide what it’s about, tagging it with topics, subfields, keywords, and SDGs. See Aboutness for the full set of subject signals and how to pick among them.

Citations get built. Each work’s reference list is extracted — from source metadata and, for open works, from the PDF — and matched to other works already in OpenAlex, producing both references and citation counts. See Works: citations and references.

This pipeline is complex and changes frequently. In the interest of openness, we share our internal monitoring dashboard publicly: view the pipeline dashboard.

Share

The same knowledge graph is available through eight channels, from no-code to whole-database. For picking among the query-time channels, see Querying; to pull the whole dataset, see Get the data.

However you access it, the data is free: everything is released under a CC0 public-domain license, with no “personal use only” carve-out and no permission to ask. Keeping it that way is a deliberate, funded plan — see What is OpenAlex’s sustainability model?