A verified, searchable record of scholarship by the current faculty and research staff of the University of Michigan School of Social Work. Explore what they study, and trust what you find.
Why this was built
This project began as a teaching tool. It grew out of trainings and workshops that show how an automated data pipeline (retrieval, verification, and semantic search) can be built and applied to a real body of scholarship.
It also does practical work for the School. Research today is scattered across dozens of databases. This brings faculty and research-staff scholarship into one place and makes it discoverable by meaning, not just keywords. Making our own work easier to find is a first step toward bringing current research into the curriculum.
3,501verified publications
70faculty & research staff
1982–2026years of scholarship
Does semantic search actually help?
Scholars describe the same idea in different words. A plain keyword search returns only the exact words you type, so it misses relevant work, especially where language is contested or evolving. Semantic search reads for meaning instead. It finds records that express your idea even when they share no words with your query. Below are real searches of this database: keyword matching against the semantic search.
You search “LGBTQ”
37 keyword matches+12 more found only by meaning
MISSED BY KEYWORD. Different words, same idea: sexual minority · gender-diverse · queer · “lesbian, gay, bisexual and transgender”
e.g. Intimate partner violence among sexual minority populations
You search “poverty”
116 keyword matches+9 more found only by meaning
MISSED BY KEYWORD. Different words, same idea: material hardship · antipoverty · income insecurity
e.g. $2.00 a Day: Living on Almost Nothing in America
You search “adolescent drug use”
45 keyword matches+14 more found only by meaning
MISSED BY KEYWORD. Different words, same idea: substance use · marijuana / alcohol use · inhalant use
e.g. Transitions in current substance use from adolescence to early-adulthood
You search “food sovereignty”
13 keyword matches+15 more found only by meaning
MISSED BY KEYWORD. Different words, same idea: food security · Indigenous food systems · “food is medicine”
e.g. Innovations in Food is Medicine Through Centering Cultural Connections and Local Food Systems
Illustrative counts from a live run of the database. In each case the extra papers never contain the words you typed, yet they are on topic. Identity terms, poverty framing, and Indigenous scholarship benefit most from reading for meaning.
How a publication reaches your screen
BEFORE YOU SEARCH: THE RECORD IS BUILT WEEKLY
THE MOMENT YOU SEARCH
1
2
3
4
5
GatherWe keep a canonical roster of current faculty and research staff and run automated searches against the major scholarly data sources every week.OpenAlexCrossrefPubMedWoSDeep BlueCVs
VerifyEvery record is tied to its DOI and checked against gold-standard registries. Metadata are messy, so several procedures resolve conflicts: deduplication, name disambiguation, and cross-source checks. An administrator reviews all corrections and monitors the automations. Nothing publishes itself.human approval required
The recordA database of real, vetted publications. Every title, author, and abstract is stored verbatim, never edited by software.3,501 records
Read & rankTwo very small AI models read your question for meaning and sort the records by relevance. They never write a word.reads, never writes
YouYou see the actual verified records, exactly as stored, with a link to a free, legal copy whenever one exists.
What gets in: the current inclusion criteria
Every publication here has a DOI: the permanent ID that journals attach to an article, a bit like a barcode for research. It lets us confirm the work really exists and match it to the right author. A publication shows up on the public site only when all three of these are true:
It has a DOI we can look up in a trusted, publicly maintained catalog of published research (OpenAlex, Crossref, PubMed, or Web of Science);
It is written by a current faculty or research-staff member, checked against the article’s own author list; and
A person on our team has reviewed and approved it.
Some scholarship doesn’t have a DOI yet; many older book chapters and reports never received one. We keep those on file and will add them over time, always explaining how each was checked. What we include, and why, will always be spelled out here.
The AI here reads. It doesn’t write.
Nothing you see on a result is written by AI. Every title, author, abstract, and citation is copied exactly from a trusted source and never rewritten.
The AI does two jobs: it reads the meaning of your search, and it puts the most relevant studies first. It cannot write sentences, so it cannot invent a study, alter an abstract, or “hallucinate.”
This is a small model, closer to a spell-checker than a chatbot. It runs on an ordinary computer, uses very little energy, and never sends your identity or reading history anywhere. Only your search text is used, and only to find matches.
Technical words, translated
DOI
A publication’s permanent ID number, like an ISBN but for articles. If two records share a DOI, they are the same paper. It is what we verify against.
Bibliographic registry
The public catalogs of published research: OpenAlex, Crossref, PubMed, and Web of Science. Publishers deposit into them directly, which makes them our ground truth.
Deep Blue
The University of Michigan’s open repository, where faculty work is archived and free to read.
Open access
A legal, free-to-read copy of an article. Whenever one exists, the result links straight to it.
Keyword search
The classic kind every library catalog does: find records containing the exact words you typed.
Semantic search
Search by meaning instead of exact words. “Teen drug use” finds “adolescent substance abuse” even though they share no words.
Embedding
A list of numbers a small model assigns to a text so that similar ideas land near each other. It is how semantic search works.
Reranker
A second small model that reads your question against each finalist’s title and abstract and puts the best answers first. The “% match” is its relevance estimate, a ranking signal rather than a probability.
“% match”
The percentage beside each result is the reranker’s estimate of how well that record answers your question, relative to the other results. Use it to compare results within one search. 94% does not mean “94% certain,” and it says nothing about the quality of the publication itself.
Common questions
It starts with a canonical roster of current faculty and research staff. Automated searches run against sources scholars already rely on: the bibliographic registries (OpenAlex, Crossref, PubMed, Web of Science), which are the systems publishers deposit into, plus the University’s Deep Blue repository, members’ CVs, and other public information. No protected organizational data.
Every Monday, the automation sweeps those sources for new work by people on the roster. New arrivals wait in a review queue. The administrator approves each record before it becomes public, reviews all corrections, and monitors the automations. Nothing publishes itself.
It means the record was matched, usually by its DOI, against an authoritative registry, and the metadata you see (title, authors, journal, year, abstract) came from that registry, not from typing. Hover the pill to see which source verified it.
Records we cannot yet verify (older book chapters, technical reports, works without DOIs) are kept privately for review, and are not shown publicly until the administrator approves them.
The strict criteria right now: a record appears publicly only if (1) it carries a DOI that resolves against a gold-standard registry (OpenAlex, Crossref, PubMed, Web of Science), (2) it is attributed to a current faculty or research staff member with that authorship confirmed by evidence, and (3) the administrator has approved it. The DOI is our foremost criterion because it points to a real-world artifact: a specific work that demonstrably exists and can be retrieved.
That standard is why some work isn't shown yet. Many book chapters, older ones especially, were never issued DOIs, and technical reports generally fall outside the peer-reviewed record. Those entries are kept privately, not discarded. Moving forward, we will incorporate non-DOI works that point to a real artifact (books with ISBNs, reports with stable repository links), each with an explicit explanation of how it was verified. The criteria will broaden, always openly and never quietly.
Google Scholar is a search engine, not a database. It crawls the open web and indexes anything that looks scholarly — journal articles, but also preprints, slide decks, syllabi, duplicate uploads, and citations to works it has never actually seen. There is no public interface for retrieving its data in bulk, no stable identifier attached to each entry, and no guarantee that two listings for the same paper are really the same record. That makes Scholar excellent for discovery and unsuitable as a source of truth.
This directory is built from registries that publishers deposit into — OpenAlex, Crossref, PubMed, Web of Science — where each work carries a DOI and structured metadata we can check. We can verify a Crossref record against its DOI; we cannot verify a Scholar listing against anything. So Scholar is a fine place for you to find a paper’s DOI, but it is the DOI, not the Scholar entry, that we ingest.
Use it if it is a stable, publicly resolvable DOI. Paste it into Verify on your My publications page and we check it against Crossref, OpenAlex, and DataCite. Registered DOIs from journals, from preprint servers (SSRN, arXiv), and from U-M Deep Blue all resolve and are accepted.
Some platforms mint internal identifiers or private links that look like DOIs but don’t resolve for anyone outside that platform, or that point to a re-hosted copy rather than the work of record. Those are not accepted, because the entire purpose of the DOI here is that anyone can follow it to the real, canonical artifact. Note that a repository Handle (hdl.handle.net/…) is not a DOI and does not count. A resolvable DOI must resolve in OpenAlex, Crossref, or DataCite.
If your work genuinely has no resolvable DOI, you can mint one by depositing it in U-M Deep Blue. New Deep Blue deposits are issued a resolvable DataCite DOI; some older legacy Deep Blue records carry only a Handle, so for those you would supply the article’s journal DOI instead.
One person: the administrator. That is the entire write path. Faculty and research staff sign in with their U-M account and can request changes: confirm a record is theirs, propose a correction, flag something that isn’t theirs, or deposit a new article with its PDF. Every request takes effect only when approved. Corrections must come with a DOI or the article PDF, so there is always something to verify against.
The public never has write access, and neither do the automated pipelines: even the weekly registry sweeps wait for human approval.
Two searches run at the same time: classic keyword matching, and semantic search. Your question and every publication have been converted into numerical fingerprints of meaning, so related ideas find each other even with no words in common. The two rankings are fused, and a reranker model reads your question against each finalist and orders them by how well they answer you.
None of these models writes a single word of what you see. They only decide which verified records to show first.
If it’s your record: sign in with your U-M account, open My publications, and submit a correction, which goes straight to the administrator. For anything else, use the Feedback page, available to signed-in U-M users only.