External APIs

ALMa pulls scholarly data from several public sources. Each has its own quotas, identifier shapes, and quirks. This page documents what ALMa fetches, how it batches, and what to expect when a source goes sideways.

Global network control

The same application-wide External network access switch is available in Health → System status and Settings → Connections → External APIs. It controls outbound scholarly APIs, Slack, email, Google Scholar, and hosted AI. Turning it off leaves local Library, maps, and local search usable, shows a persistent warning, and stops requests at the shared transport before any socket is opened. ALMA_DISABLE_NETWORK=1 is the operations hard override: it can force access off but cannot force a user's stored off choice on.

Find & Add remains graceful when OpenAlex's paid search pool is drained: paper search omits only the OpenAlex lane and continues Semantic Scholar, Crossref, arXiv, bioRxiv, and Europe PMC; author search falls back to matching local OpenAlex-linked identities. Import review and Health repair cards use the same credit forecast and disable work that the known remaining pool cannot cover.

OpenAlex (primary)

OpenAlex is ALMa's primary source for metadata, citations, topics, institutions, and the works graph.

• Endpoints used: /works, /works/{id}, /authors, /authors/{id}. Topics, sources, and institutions arrive embedded via select= on /works and /authors — they are not separate calls. Related works come from /works/{id} with select=related_works, then a bulk-by-ID fetch of those IDs.
• No semantic /find/works: the natural-language /find/works semantic call was removed — it 404s and is a separate paid product (~1000 credits/call). Hybrid search is now purely the lexical /works?search= path.
• API key — REQUIRED (since 2026-02-13): every request needs OPENALEX_API_KEY. OpenAlex retired the email "polite pool"; without a key you get 100 free credits/day and then HTTP 409. A free key (openalex.org/settings/api) gives standard limits — 100,000 credits/day (singleton GETs cost 0 credits, list requests 1 each), at a typical ~10 req/s. Set it in .env or via Settings → Connections → OpenAlex.
• Contact email (optional): OPENALEX_EMAIL no longer affects rate limits (the polite pool is gone) but still sets a courteous User-Agent and feeds the Crossref polite pool.
• Field projection: ALMa always sends a select= parameter so we only fetch the fields we use. The select list lives in alma/openalex/client.py::_WORKS_SELECT_FIELDS. Adding a downstream consumer of a new field requires updating the select.
• Pricing is per request class, and singletons are FREE. This is the fact that decides when to batch:

operation	cost / 1 000 calls
Get singleton (one entity by ID or DOI)	free
List + filter	$0.10
Search (full-text)	$1.00
Content download	$10.00

A free key is $1.00/day — unlimited singletons, ~10 000 list+filter, ~1 000 search. Source of truth in-repo: alma/openalex/http.py::_CLASS_COSTS_USD (+ cost_class()), which mirrors https://developers.openalex.org/api-reference/authentication. Read those constants rather than copying prices, which drift. * Batch on latency, not by rule. Because singleton GETs cost nothing, mechanically replacing them with pipe-filter calls makes things more expensive, not less. The trade only pays on a hot path: 100 singletons is ~10 s of round trips at $0, whereas one pipe-filter is ~0.3 s at $0.0001 (0.01% of the daily budget). So ALMa batches the interactive per-candidate author-enrichment loop in openalex/client.py and deliberately leaves the one-off singleton lookups alone. * Bulk-by-ID batching: the corpus rehydrator fetches metadata for many papers in one call via filter=openalex_id:W1|W2|… — up to 100 IDs per filter (the documented OR ceiling) with per-page=100, splitting on HTTP 400/414 if a URL ever runs long. Worth 1 list credit instead of N free singletons because the backlog is latency-bound. The client uses thread-local sessions so the rehydration pipeline can fan these out concurrently. * Rate limits: ALMa tracks every response's x-ratelimit-* headers and exposes them at Settings → External APIs → OpenAlex usage. If you see "no calls yet", you haven't hit the API yet — it does not mean we failed to record.

Semantic Scholar

Semantic Scholar covers what OpenAlex doesn't: pre-computed SPECTER2 vectors, paper-level recommendations, and the reference/citation graph.

• Endpoints used:
• /paper/batch for bulk metadata + specter_v2 vectors.
• /paper/search/bulk for the recent Feed/monitor ingestion lane; /paper/search for interactive/general Discovery.
• /paper/search/match for title→paper identity (the corpus title-resolution sweep and author-identity triangulation).
• /paper/{id}/references and /paper/{id}/citations for the graph lane.
• /recommendations/v1/papers/forpaper/{id} (single seed, all-time) and POST /recommendations/v1/papers (positive + negative seeds, new work only) for the s2_related recommendation channel — a different host path than the rest of the graph API.
• /author/batch, /author/search, and /author/{id}/papers for author identity resolution.

Documented limits (from the live graph/v1/swagger.json)

endpoint	limits
POST /paper/batch	500 lookup ids, 10 MB response, 9 999 citations
POST /author/batch	1 000 author ids, 10 MB
GET /paper/search	limit ≤ 100, ≤ 1 000 relevance-ranked results total, 10 MB
GET /paper/search/bulk	1 000 rows/call, no limit/offset parameter, token pagination, 10 M total, no nested data
GET /paper/search/match	one closest-title row; 404 on a miss
GET /paper/{id}/citations · /references	limit ≤ 1 000
GET /author/{id}/papers · /author/search	limit ≤ 1 000

/paper/{paper_id} and its sub-resources accept DOI:<doi>, CorpusId:<id>, ARXIV:<id>, PMID:, PMCID: — never pre-resolve an identifier to a paperId with an extra request.

• Field projection is per-endpoint, not global (ENDPOINT_FIELDS + project_fields() in discovery/semantic_scholar.py). S2 does not accept the same fields everywhere, and an unsupported field is a hard HTTP 400, not a silent drop. Verified live 2026-07-27:

endpoint	tldr + embedding.specter_v2
POST /paper/batch	accepted
GET /paper/search · /paper/search/match	accepted
GET /paper/search/bulk	rejected (400)
POST /recommendations/v1/papers	rejected (400)
GET /recommendations/v1/papers/forpaper/{id}	rejected (400)

The registry is additive — an endpoint declares what it allows, and an unregistered endpoint raises rather than inheriting the global list. The previous subtractive patch (one _BULK_UNSUPPORTED_FIELDS constant) failed open: POST /recommendations inherited the two rejected fields and returned HTTP 400 on every call for the life of the lane, hidden behind if resp.status_code != 200: return []. Non-429 4xx now logs at WARNING.

The trap that makes a registry necessary rather than a probe: the field error only fires when the query actually matches something. GET /forpaper with the full field set returns a clean 200 {"recommendedPapers": []} on from=recent (empty for older seeds) and a hard 400 on from=all-cs. An endpoint can look field-compatible right up until it starts returning results. * Bulk search is two-step (bulk → batch): bulk cannot return tldr or the SPECTER2 vector, so ALMa requests the bulk-supported subset for breadth, reranks locally, then issues a single /paper/batch to hydrate only the slice it keeps. * Bulk has no relevance ordering. Its sort accepts only paperId, publicationDate and citationCount, and defaults to paperId:asc — a sha hash. Left at the default this lane returned the 1 000 lowest-paperId matches out of millions, deterministically, so a keyword monitor surfaced the same papers forever. ALMa now sends sort=publicationDate:desc (the pool means "newest matching work") and reconstructs relevance locally with core.scoring_math.query_match_score. Note the spec's caveat: records with an undefined sort value sort last, so date-less papers fall outside the 1 000-row window — accepted for this recent-ingestion lane only. * year is pushed server-side (year=2020-). It used to be filtered in Python after downloading all 1 000 rows. * API key — strongly recommended: set SEMANTIC_SCHOLAR_API_KEY (free at semanticscholar.org/product/api), in .env or via Settings → Connections → Semantic Scholar. Without one you share the anonymous worldwide pool (~5,000 requests / 5 min across all anonymous clients), so 429s are frequent — they're the root cause of the multi-minute Discovery graph-lane stalls. With a key you get a dedicated ~1 RPS. * /paper/batch has TWO independent budgets, and conflating them is a bug: * ID budget — 500 lookup ids per call. One paper contributes up to two (S2 ID + DOI), so a "500 papers" setting emits 1 000 ids. * Payload budget — 10 MB per response. Measured 2026-07-27 with the full field set: 18 899 bytes/row, dominated by the 768-d SPECTER2 vector. 500 vector-bearing rows ≈ 9.4 MB, i.e. 94% of the hard cap. * plan_paper_batch() sizes each request from the requested field projection (a narrow projection legitimately earns a bigger batch), keeps a 75% safety margin, and recursively splits on a size refusal. The batch knob surfaced in Settings reads that same plan via services/eta.py, so the estimate and the run can no longer disagree. * Results preserve the request order by lookup id (DOI / S2 ID / OpenAlex ID). * Compacting the response shifts good papers onto bad IDs and corrupts state. ALMa preserves the original index. * null rows in the middle of a response are real — the lookup didn't match. * One batched-vector primitive: fetch_vectors_for_identifiers() returns an IdentifierFetchOutcome (rows keyed by requested id, plus disjoint terminal_ids / retryable_ids). fetch_papers_batch and services/s2_vectors are both thin callers — a second resilient-split implementation is exactly what CLAUDE.md → "fix the primitive" forbids. * Recommendation pools: GET /forpaper/{id} defaults to from=recent, which is empty for older seeds (measured: 0/20 results on a 2017 seed vs 20/20 with from=all-cs). ALMa always sends from=all-cs — that lane exists for foundational/older related work. The POST endpoint has no from parameter at all and only returns 2025–26 work, so it is a frontier source, not a general related-papers source. * Author hydration batches. /author/batch takes 1 000 ids; the sweep gathers a whole phase's S2 ids into one request instead of one request per author (at 1 RPS that was one second per author). * Failure classification: * Retryable failures (429, 5xx, network) stay retryable — they don't become terminal "no match". * Only validation failures (4xx other than 429) split down to singleton lookups and mark only those papers as lookup_error. * search_papers(raise_on_rate_limit=True) surfaces a 429 to its caller as SemanticScholarBatchError(status_code=429) so the title-search rescue can defer instead of stamping terminal_no_match. The legacy default (silent empty list) is preserved for non-critical callers like interactive search. * Adaptive throttle: any 429 observed by the shared HTTP client (core/http_sources.SourceHttpClient) engages a 10-second floor on the per-request interval for the next 30 seconds. Retries: 5 attempts, jittered exponential backoff capped at 60 seconds. Fresh 429s within the cooldown re-arm the floor. While the cooldown is armed, Discovery and Feed drop the S2 source for the rest of that refresh pass (via is_in_adaptive_cooldown()) instead of having each lane queue behind the floor and wait out its deadline; the window self-clears after 30s (discovery/source_search.py). * Terminal statuses for vector fetch: unmatched, missing_vector, lookup_error, bad_local_doi. bad_local_doi is set before any HTTP call when the local DOI fails the registry-shape regex ^10\.\d{4,9}/.+; it never reaches S2. Terminally-missed papers stay eligible for explicit local SPECTER2 compute. The trigger papers_clear_fetch_status_on_id_change (see api/deps.py:init_db_schema) drops these terminal rows whenever papers.doi or papers.semantic_scholar_id actually changes — so a paper hydration step that finds a better DOI re-enters the fetch pool automatically. * Title-search rescue: papers that miss /paper/batch get one /paper/search/match call each — the purpose-built closest-title endpoint, which returns one row instead of N ranked ones and still carries abstract + the SPECTER2 vector. This lives in services/title_resolution.py (not s2_vectors.py, which delegates to it), capped at TITLE_RESOLUTION_PER_RUN_BUDGET = 500 papers per run. The first 429 short-circuits the rest of the batch's rescue.

Three contract details that make it not a drop-in for /paper/search: the envelope is {"data": [row]}; a miss is HTTP 404 (Title match not found), which is a normal outcome, not an error; and matchScore is unbounded (131.8 on an exact title) and cannot be requested in fields. Acceptance therefore stays with ALMa's local Jaccard 0.92 / |Δyear| ≤ 1 contract — never matchScore.

match_paper_by_title() returns the raw S2 row so externalIds and authors[].authorId survive. application/author_identity needs both, and ALMa's normalized candidate shape drops them: it flattens authors to a comma-joined string, so the previous search_papers-based code raised AttributeError on a.get("authorId") and never extracted a preprint id. * Identity resolution (OpenAlex-first): the Resolve missing identity sweep (services/title_resolution.py) handles title-only papers with no usable identifier. It tries OpenAlex /works?search first (the larger, cheaper, higher-throughput pool) and only falls back to S2 /paper/search (1 RPS) when OpenAlex misses — so an OpenAlex-cold corpus can't blow the S2 quota. An accepted OpenAlex match fills the full work from that same search response (no second fetch), while the S2 fallback captures the SPECTER2 vector it carries. Same Jaccard 0.92 / |Δyear|≤1 contract as the rescue above. * DOI hygiene (core.utils.canonical_lookup_doi): DOIs sent to S2 are lowercased, URL-decoded, and stripped of trailing publisher fragments (/pdf, /full, /abstract, /epdf, /meta). The match-side bookkeeping uses the same canonical form so case-only differences round-trip cleanly.

Crossref

Crossref is the DOI authority and a metadata fallback when OpenAlex doesn't have a paper.

• Endpoints used:
• /works/{doi} — singleton DOI lookup; used by per-paper _hydrate_via_crossref and other one-off paths.
• /works?filter=doi:DOI1,doi:DOI2,...&rows=50 — batched DOI lookup via discovery.crossref.fetch_works_by_dois. Phase 2 of the bulk corpus rehydrator uses this to resolve up to 50 DOIs per HTTP call (~50× round-trip reduction at full backlog vs the singleton path).
• Polite pool: Set CROSSREF_MAILTO to identify yourself. They ask for it; honour the request. Crossref retuned its REST limits on 2025-12-01: the list/search path — what ALMa hits via /works?query — is now the stricter ceiling at 3 req/s polite / 1 req/s anonymous. Single-record /works/{doi} is looser (10 req/s polite), but one client serves both paths, so ALMa paces to the search ceiling: 0.34s interval polite / 1.05s anonymous, with concurrency 3 polite / 1 anonymous (core/http_sources.py).
• Used as a fallback, not the primary path. Most papers resolve through OpenAlex first.

Europe PMC

Europe PMC indexes PubMed/MEDLINE, PMC, preprint servers and Agricola — the life-sciences counterpart to S2's CS-leaning corpus. Free, no key, abstracts inline.

• Endpoint used: GET /search with format=json and resultType=core. core is the only tier that returns abstractText; lite is cheaper but every candidate would arrive unscoreable.
• Filters are pushed server-side into the query string — FIRST_PDATE:[YYYY-01-01 TO 3000-12-31] for a year floor and OPEN_ACCESS:y for OA-only.
• Response shape gotcha: the journal title is nested at journalInfo.journal.title, not a top-level journalTitle. authorString is already display-ready.
• Not a metadata authority. OpenAlex and Crossref remain the identifier/metadata sources of truth; this adapter contributes discovery candidates plus the DOI / PMID / PMCID needed to reconcile them.
• Paced at ~5 req/s, 2 concurrent (core/http_sources.py). EBI publishes no hard per-second figure and asks only that clients be reasonable.

arXiv and bioRxiv

• arXiv: ALMa uses arXiv's metadata API to resolve preprints not yet indexed by OpenAlex.
• bioRxiv (also covers medRxiv): same fall-through pattern, but bioRxiv has no keyword-search endpoint (date-range / DOI / category only). ALMa pulls a shared recent date window, caches it per (server, interval) for 300s so every keyword monitor/lane in one refresh shares a single network pull, then filters and re-ranks locally per query (discovery/biorxiv.py).
• Abstract recovery (task 05): both servers' structured APIs also back-fill a missing abstract on a paywalled paper that has an arXiv/bioRxiv preprint twin — arXiv's Atom summary (by id, or a title-field ti:"…" search) and bioRxiv's /details/{server}/{doi} abstract. This runs ahead of any landing-page HTML scrape (cleaner, paywall-free), is fill-only, and stamps a per-source ledger reason (services/preprint_abstract.py). bioRxiv is direct-DOI/URL only — it has no keyword endpoint, so the title-search fallback is arXiv-specific.

These are read-only fall-through paths. Each has its own DOI prefix that triggers the preprint↔journal dedup engine:

DOI prefix	Source
10.48550/arXiv.*	arXiv
10.1101/*	bioRxiv / medRxiv
10.31234/*	psyArxiv (OSF)
10.31219/*	OSF
10.26434/chemrxiv*	chemRxiv
10.20944/preprints*	MDPI Preprints

Google Scholar (scholarly)

The scholarly package is opt-in and used only for author identity resolution as a tiebreaker — never for primary metadata fetches.

• When it's used: when OpenAlex / S2 don't disambiguate an author and you've added them by name.
• When it's not used: anything that can be answered by OpenAlex. You can leave scholarly uninstalled and most flows still work.
• Stability warning: Scholar has no public API. The library scrapes; expect occasional rate-limiting / breakage when Google changes the page. Set SCHOLAR_RETRY_DELAYS to tune backoff.

OpenAI

OpenAI is optional and currently used as an embedding provider. Configure it in Settings → Intelligence → AI provider and store the key in .env or the secret store. See AI capabilities.

Slack

Slack Web API, bot-token based.

• Token: SLACK_TOKEN, or the secret store (slack.bot_token).
• Outbound channel: SLACK_CHANNEL, or slack_channel in settings.
• Inbound (capture) channel: slack_inbox_channel, Settings only.

ALMa uses slack-sdk through ONE client, alma.slack.client.SlackNotifier: alerts post with chat.postMessage (Block Kit), capture reads with conversations.history and acknowledges with reactions.add. No webhooks, and no second client — see External integrations.

Identifier reference

Common ID shapes you'll see across the API:

Shape	Source	Example
OpenAlex Work ID	OpenAlex	W2123456789
OpenAlex Author ID	OpenAlex	A2123456789
Semantic Scholar paper ID	S2	649def34f8be52c8b66281af98ae884c09aef38b
Semantic Scholar corpus ID	S2	12345678
DOI	Crossref / authority	10.1038/nature12373
ORCID	ORCID	0000-0002-1825-0097
arXiv ID	arXiv	2401.12345

All five are stored on the relevant rows when known. Identity resolution (authors and papers) tries to fill in as many as possible.