OAcquire

A Multi-Source Corpus Acquisition Layer for Agentic Systematic Reviews

🌐 Try it in your browser — no installation: openacquire.streamlit.app

---

Screening a systematic review is a solved problem. Getting the full texts is not.

Between a completed search (Scopus, OpenAlex, PubMed) and any downstream analysis — manual synthesis, text mining, or an LLM ingestion pipeline — sits an unglamorous, error-prone, and almost entirely undocumented step: turning a list of identifiers into a directory of validated full-text PDFs. In practice this is done by hand, by ad-hoc scripts, or by tools that quietly bypass paywalls. None of these produce a record you can report in a PRISMA flow diagram.

OAcquire is the acquisition layer for that gap. It resolves DOIs, PMIDs, and PMCIDs to legally retrievable open-access full texts by cascading across seven independent sources, validating every candidate document, and emitting an auditable retrieval log. It is designed to be called three ways: by a human at a command line, by a non-technical researcher in a browser, and — the case it was actually built for — by an autonomous agent as a single typed function call.

Why this exists

Existing tools fail research workflows in one of three ways:

	Legal	Multi-source	Auditable log	Agent-callable
Manual download	✅	—	❌	❌
Reference-manager "find PDF"	✅	Partial	❌	❌
Scraping tools with Sci-Hub backends	❌	✅	❌	Partial
OAcquire	✅	✅ (7 sources)	✅	✅

OAcquire does not and will not circumvent paywalls. Coverage is bounded by what is legitimately open access — and quantifying exactly where that boundary falls, per discipline, is one of the things the tool is designed to measure.

Design

Three layers, deliberately separated so each is usable and testable in isolation:

LAYER 3  cli.py       Batch CSV workflows. JSON summary → stdout, logs → stderr.
LAYER 2  core.py      retrieve_pdf() — one typed call, no side effects, no sys.exit.
LAYER 1  sources.py   Seven source adapters + the candidate cascade.

The cascade is the core contribution. Naïve retrievers ask each source for a URL and give up when it fails. OAcquire collects every candidate URL from every enabled source — all OA locations from Unpaywall, all link objects from Crossref, every anchor and citation_pdf_url meta-tag on the landing page — deduplicates them, and then attempts each in priority order. Every attempt is validated: magic-byte inspection rejects HTML served with a PDF content-type, and a size floor rejects publisher stub pages. Failures are recorded with a diagnostic code, not silently dropped.

Sources

Attempted in this order:

#	Source	Requires	Notes
1	PubMed Central	—	NCBI OA API; auto-converts PMID/DOI → PMCID
2	Europe PMC	—	EBI REST API; strong biomedical coverage
3	Unpaywall	--email	Per their Terms of Service
4	OpenAlex	--openalex-api-key	Opt-in via --openalex
5	Semantic Scholar	—	Opt-in via --semantic-scholar; strong for CS/engineering
6	Crossref	--email (optional)	Publisher metadata link objects; ≤3 workers
7	Landing page	—	citation_pdf_url meta-tags and anchor heuristics

Install

pip install oacquire

Or from source:

git clone https://github.com/pecesama/oacquire.git
cd oacquire
pip install -e ".[dev]"

Quick start

oacquire --input examples/literature.csv --email you@university.edu

PDFs land in pdfs/; the audit log in download_report.csv; a machine-readable summary on stdout.

The input CSV uses a minimal, source-agnostic schema compatible with exports from any database or reference manager:

Column	Required	Notes
Title	Yes	Used as filename fallback
Authors	Yes	Author list
DOI	Yes*	Primary identifier
PMID	No	Enables PMC / Europe PMC lookup
PMCID	No	Skips ID conversion

* At least one of DOI, PMID, or PMCID per row.

Agent / programmatic use

retrieve_pdf() is the interface OAcquire was built around. It returns a typed DownloadResult — no prints to stdout, no sys.exit, no exceptions on the expected failure paths.

from oacquire import retrieve_pdf, build_session

session = build_session()   # reuse across calls

result = retrieve_pdf(
    doi="10.3390/sym17122083",
    email="you@university.edu",
    session=session,
)

if result.ok:
    ingest_to_vector_store(result.pdf_path)
elif result.status == "not_found":
    flag_for_manual_screening(result.doi, result.note)

Field	Type	Description
ok	bool	True if a validated PDF was saved
status	str	downloaded · not_found · failed · skipped
source	str	Which source succeeded
pdf_path	Path | None	Path to the saved file
pdf_url	str	URL used
note	str	Diagnostic code (see below)

Agent loop

OAcquire was built for exactly this pattern: an agent resolving a list of screened works into a full-text corpus, routing each outcome without human intervention. Because retrieve_pdf() never raises on the expected failure paths and returns a typed result, the loop stays branch-clean.

from oacquire import retrieve_pdf, build_session

session = build_session()   # reuse one session across the whole corpus

# e.g. the output of a systematic-review screening stage, or an
# OpenAlex / arXiv query feeding a RAG pipeline
screened = [
    {"doi": "10.3390/virtualworlds4040056", "title": "Extended Reality in CS Education"},
    {"doi": "10.3390/app15158679",          "title": "ML and Generative AI in Learning Analytics"},
    # ...
]

corpus, unresolved = [], []
for work in screened:
    result = retrieve_pdf(**work, email="you@university.edu", session=session)
    if result.ok:
        corpus.append(result.pdf_path)          # → chunk, embed, index
    elif result.status == "not_found":
        unresolved.append((work["doi"], result.note))   # → PRISMA exclusion log
    else:
        unresolved.append((work["doi"], f"{result.source}:{result.note}"))

print(f"{len(corpus)} full texts acquired; {len(unresolved)} unresolved")

The unresolved list is not a dead end — it is your PRISMA "reports not retrieved" count, already itemised with the reason each work failed. That is the number reviewers ask for and that most pipelines cannot produce.

The audit log

download_report.csv is not a convenience — it is the point. Each row records what was attempted, what succeeded, and why a failure occurred:

note value	Meaning
downloaded	Validated PDF saved
already_exists	Skipped (--skip-existing)
pdf_too_small:{N}_bytes	Stub / placeholder PDF rejected
not_pdf_magic_bytes:{ct}	HTML served as PDF — rejected
http_error:403	Paywall or bot block
all_candidates_failed:{src}:{why}	Every candidate URL exhausted

This distinguishes "no open-access version exists" from "an OA version exists but retrieval failed" — a distinction that matters for honest PRISMA reporting and that, to our knowledge, no comparable tool records.

CLI reference

--input                     CSV input (required)
--outdir                    PDF output directory (default: pdfs)
--report                    Audit log path (default: download_report.csv)
--email                     Required for Unpaywall; enables Crossref polite pool
--openalex                  Enable OpenAlex (requires --openalex-api-key)
--openalex-api-key          Free key at openalex.org
--semantic-scholar          Enable Semantic Scholar
--semantic-scholar-api-key  Optional; raises rate limits
--doi-column / --pmid-column / --pmcid-column / --title-column
                            Override auto-detected column names
--workers                   Parallel threads (default 1; keep ≤3 with Crossref)
--skip-existing             Resume an interrupted run
--max / --delay / --timeout Batch and politeness controls
--no-pmc / --no-europepmc / --no-unpaywall / --no-crossref / --no-landing-page
                            Disable individual sources (used for ablation studies)
--verbose                   Per-record progress to stderr
--version / --citation      Version string / canonical citation

Web interface

For researchers who don't use a command line, the full pipeline is available in the browser at openacquire.streamlit.app — upload a CSV or paste DOIs, toggle sources, watch progress live, and download the PDFs as a ZIP plus the audit log as CSV.

To run it locally:

pip install -e ".[app]"
streamlit run app/streamlit_app.py

Responsible use

• No paywall circumvention. OAcquire locates only legally open versions: OA journals, preprint servers, institutional repositories, and publisher-authorised links. It has no Sci-Hub backend and will not accept one.
• Polite traffic. Requests are delayed and rate-limited by default, with exponential backoff. Keep --workers ≤ 3 when Crossref is enabled.
• Identify yourself. --email is required by Unpaywall's terms and places you in Crossref's polite pool.

Citing OAcquire

If OAcquire contributes to work you publish, please cite the paper, not the software archive — this keeps citations consolidated in indexed databases:

@article{santanamancilla2026oacquire,
  title   = {OAcquire: A Multi-Source Corpus Acquisition Layer for Agentic Systematic Reviews},
  author  = {Santana-Mancilla, Pedro C.},
  journal = {arXiv preprint},
  year    = {2026}
}

For exact reproducibility, additionally cite the version you ran (e.g. "retrieval performed with OAcquire v2.0.0, DOI 10.5281/zenodo.21367807"). Run oacquire --citation to print the current canonical citation.

Contributing

Issues and pull requests are welcome — see CONTRIBUTING.md. New source adapters are especially valuable; the adapter contract is a single function returning (url, note) tuples, and sources.py documents it.

AI usage disclosure

Portions of this codebase and its documentation were developed with the assistance of generative AI tools (Anthropic Claude). All architectural decisions, the source-cascade design, the validation strategy, and the evaluation protocol were made by the human author, who reviewed, tested, and validated all AI-assisted output and is solely responsible for its correctness.

👨‍💻 Author

Developed by Pedro C. Santana-Mancilla — pedrosantana.mx
As part of his own automation and research tools efforts at the IHCLab Research Group School of Telematics, Universidad de Colima.

MIT licensed.