litfetch

Resolve a scholarly article identifier to its retrievable artifacts — the full-text body and any supplementary material — and fetch their bytes.

litfetch is two cooperating seams:

• a fetch ladder — pluggable Fetcher backends (PMC Open Access S3, Europe PMC, Elsevier OA) tried in priority order; the first to serve the body wins, returning a Blob (a File plus its bytes);
• an optional resolver layer — pluggable Resolvers that enrich what you know about a paper (pmid → pmcid/doi, etc.) so the ladder can act.

You hand it an ArticleIds bundle (any of pmid / pmcid / doi). Resolution is demand-driven: a resolver only runs when the next fetcher needs an identifier you don't yet have, and runs at most once.

An article is modelled as a file-set: a collection of File references (the body in its various media types, plus supplementary material, distinguished by FileKind), each hosted upstream. litfetch fetches the raw artifacts and reports their access terms; it does not render them. To turn a fetched JATS/Elsevier body into markdown, run litdown on the bytes yourself (see Render to markdown).

The examples below are a tour; docs/api.md is the full reference for the public surface.

Install

pip install litfetch

bioRxiv / medRxiv preprint full text needs a browser-fingerprint HTTP client, enabled by the biorxiv extra:

pip install 'litfetch[biorxiv]'

Usage

Fetch the body

Hand fetch_body an ArticleIds; the default ladder serves the first available body as a Blob:

from litfetch import ArticleIds, fetch_body

blob = await fetch_body(ArticleIds(pmcid='PMC5334499'))
if blob:
    print(blob.file.source, blob.file.media_type, len(blob.content))

Render to markdown

litfetch returns raw bytes, not markdown. Convert a JATS/Elsevier body with litdown — you pick and pin the converter:

import io
import litdown
from litfetch import ArticleIds, fetch_body

blob = await fetch_body(ArticleIds(pmcid='PMC5334499'))
if blob:
    markdown = litdown.convert(io.BytesIO(blob.content))

Inject your own resolver

A resolver is an async (ArticleIds, Http) -> ArticleIds — the session running it supplies the Http. Enrich from whatever you have — a corpus client, a local cache, an API — and merge it in (this one ignores Http, hence _http):

from litfetch import ArticleIds, Http, fetch_body

async def my_resolver(ids: ArticleIds, _http: Http) -> ArticleIds:
    if not ids.pmid:
        return ids
    pmcid, doi = await my_corpus.lookup(ids.pmid)
    return ids.merge(ArticleIds(pmcid=pmcid, doi=doi))

blob = await fetch_body(ArticleIds(pmid='29622564'), resolver=my_resolver)

Use a bundled resolver

Bundled resolvers are constructed with their config, then passed in the same slot. chain(...) composes several (yours first, fallbacks after); it stops once every identifier is known:

from litfetch import ArticleIds, fetch_body
from litfetch.resolvers import SemanticScholarResolver, NcbiIdConverterResolver, chain

resolver = chain(
    my_resolver,                              # your own
    SemanticScholarResolver(api_key=S2_KEY),  # bundled
    NcbiIdConverterResolver(tool='myapp'),    # bundled
)
blob = await fetch_body(ArticleIds(pmid='29622564'), resolver=resolver)

Polite-pool identification (NCBI/Crossref email, Unpaywall's required email) comes from a session contact, not a hardcoded default — set it on the session: async with litfetch.Session(contact='you@example.org') as s: await s.fetch_body(...).

default_resolver() is a batteries-included, keyless chain (Europe PMC search + NCBI ID Converter).

No resolver — you already hold the IDs

A non-PubMed paper you only have a DOI for, plus your own Elsevier key:

blob = await fetch_body(
    ArticleIds(doi='10.1016/j.cell.2020.01.001'),
    credentials={'elsevier_api_key': key},
)

Supplementary material

list_files enumerates the file-set (references, no bytes); fetch_file materialises one:

from litfetch import ArticleIds, FileKind, list_files, fetch_file

files = await list_files(ArticleIds(pmcid='PMC5334499'), kind=FileKind.SUPPLEMENTARY)
for file in files:
    blob = await fetch_file(file)

Access terms

Read the licence from the fetched bytes, falling back to an access authority (Unpaywall) when the bytes carry none:

from litfetch import extract_source_metadata, resolve_access

meta = extract_source_metadata(blob)          # from the JATS/Elsevier bytes
if meta.licence is None:
    meta = await resolve_access(ArticleIds(doi='10.1016/j.cell.2020.01.001'))

Resolvers stand alone

Each resolver is usable on its own as a cross-reference tool, independent of fetching. A resolver is given the Http to use, so run it inside a session:

from litfetch import ArticleIds, Session
from litfetch.resolvers import SemanticScholarResolver

async with Session() as s:
    ids = await SemanticScholarResolver()(ArticleIds(doi='10.1016/j.cell.2020.01.001'), s)
print(ids.pmid, ids.pmcid)

Batch: one session, a scope per paper

The one-shot functions above each open a throwaway session. For many papers, hold one Session (pooled connection, shared pacing) and open a scope per paper — the scope caches within itself, so a duplicate upstream call (e.g. Unpaywall for both licence and PDF) is fetched once:

from litfetch import ArticleIds, Session

async with Session() as session:
    for pmid in pmids:
        async with session.scope() as s:
            blob = await s.fetch_body(ArticleIds(pmid=pmid))
            access = await s.resolve_access(ArticleIds(pmid=pmid))

Extending

• A new body fetcher: implement the Fetcher protocol — a name, a requires: frozenset[str] of the ArticleIds fields it needs, and an async fetch(ids, *, credentials, http) returning a body Blob or None. Add it to a fetchers= list (or your own default_fetchers).
• A new file source: implement the FileSource protocol — a name, and async list_files(ids, ...) / fetch_file(file, ...) — to enumerate and materialise an article's file-set (body renditions and supplementary alike).
• A new resolver: write an async ArticleIds -> ArticleIds that fills gaps via ArticleIds.merge and never overwrites a known id.

Development

uv sync
uv run ruff check . && uv run ruff format --check .
uv run pyright
uv run pytest