@rdf-label-cache/client

A really thin (one file, zero dependencies), isomorphic client for the RDF Label Cache. It is the canonical example of how to consume the service well — it encodes, in code, the guidance from consuming.md and FAQ.md:

• Parallel by default, bounded to ~100 in flight — the per-connection H2/H3 stream limit. Independent, edge-cached GETs multiplexed over one warm connection.
• Prefers the browser cache. It never busts it (no cache: "no-store"), so a returning view is served from the browser's own HTTP cache at 0 bytes on the wire. In the browser it adds only an in-flight de-dupe; outside the browser (Node/Workers/Deno — no shared HTTP cache) it keeps a small memo LRU.
• Client-side language fallback. There is no server-side fallback; a ?lang= miss is corrected with a cheap, separately-cached second call.
• Warms the connection with preconnect() (browsers).

Install

pnpm add @rdf-label-cache/client

Use

import { createLabelClient } from "@rdf-label-cache/client";

const labels = createLabelClient({
  base: "https://label-cache-orders.<subdomain>.workers.dev",
});

// Browser: warm the TLS/QUIC handshake as the page loads.
labels.preconnect();

// One label (untagged). Applies predicate order + language fallback.
await labels.resolve("https://schema.org/name");            // → "name"

// A specific language, falling back to the untagged label on a miss.
await labels.resolve("https://schema.org/name", { lang: "en" });

// Many at once — deduped, parallel, bounded. This is the hot path.
await labels.resolveMany([
  "https://schema.org/Person",
  "https://schema.org/jobTitle",
  "http://www.w3.org/2004/02/skos/core#Concept",
]);
// → { "https://schema.org/Person": "Person", … }

Need the raw JSON-LD instead of a single string? Use document(iri, lang?) / documentMany(iris, lang?) and pick with the exported pickLabel.

Getting real RDF out (still zero-dep)

The raw response is JSON-LD — as delivered it is only valid JSON; it becomes valid RDF once interpreted against the context (so prefLabel → skos:prefLabel, the "en" key → a language tag). You don't need a JSON-LD processor for that: the service's context is a trivial subset (prefix maps + @container: @language), so toQuads expands faithfully, driven by the service's own fetched context — correct even for self-hosters with custom predicates.

// Zero-dep: plain RDF/JS-shaped { subject, predicate, object, graph } objects.
const triples = await labels.toQuads(await labels.document("http://…/Concept", "en"));

// Already using n3? Hand it the DataFactory to get n3's native quads, straight
// into a Store — no jsonld dependency, no reinvention:
import { DataFactory, Store } from "n3";
const store = new Store();
store.addQuads(await labels.toQuads(doc, { factory: DataFactory }));

@none expands to an xsd:string literal, a real tag to an rdf:langString. toQuadsMany(docs, opts?) concatenates across documents.

Prefer a full JSON-LD processor?

This library intentionally does not depend on jsonld or n3 — most label cache consumers already have them, and dependency injection beats a bundled (or even optional) dependency. Two supported routes:

1. toQuads({ factory }) as above — enough for this service's data, zero-dep.
2. Bring your own parser for full JSON-LD 1.1 / streaming. Feed the raw response to jsonld-streaming-parser (the n3 companion) or jsonld, with the context served from memory via getContext() (it's versioned and immutable — cache it once, per consuming.md §3), so no per-document context round trip.

API

| Member | Purpose | |---|---| | createLabelClient(opts) / new LabelClient(opts) | Construct once, reuse — keeps the connection warm and caches live. | | .resolve(iri, { lang?, fallback? }) | IRI → label string (or null). | | .resolveMany(iris, opts?) | IRIs → { [iri]: label \| null }, concurrent & deduped. | | .document(iri, lang?) / .documentMany(iris, lang?) | Raw JSON-LD (or null on a miss). | | .toQuads(doc, { factory? }) / .toQuadsMany(docs, …) | Expand JSON-LD → RDF triples (zero-dep, or your RDF/JS DataFactory). | | .getContext() | The versioned, immutable context document (cached) — for a BYO JSON-LD parser's documentLoader. | | .url(iri, lang?) | The request URL for a pair (for <link rel=preload>, debugging). | | .preconnect() | Browser: preconnect to the origin. No-op elsewhere. | | pickLabel(doc, order?, lang?) / pickFromMap(map, lang) | Pure pickers, usable without a client. | | lruStore(max?) | The default memo store; pass your own via cache. |

Options

createLabelClient({
  base,                 // required: your deployed Worker origin
  order,                // preference by term ALIAS (doc keys, not IRIs); default ["prefLabel","label","title","name"]
  concurrency,          // max in-flight; default 100
  cache,                // "auto" (default) | true | false | your own LabelStore
  fetch,                // injectable fetch; default globalThis.fetch
});

cache: "auto" turns the memo LRU off in browsers (the HTTP cache already does this job and a hand-rolled map mostly duplicates it) and on everywhere else (server callers have no shared HTTP cache). Only successful documents are memoized — a miss is never cached, so a label seeded shortly after a 404 shows up on the next call.

order is matched against the term aliases in the response (the object's keys, e.g. prefLabel), not IRIs. The default matches the aliases this service's context emits. Because a term keeps its alias even if a deployment maps it to a different IRI, remapping (say name → foaf:name) needs no change; only a context that renames the aliases does — then pass your own order.

Why so thin?

Because the service is designed so the client can be thin: the edge cache, HPACK/QPACK header compression, and HTTP/3 stream multiplexing do the heavy lifting. The client's whole job is to fire requests in parallel, get out of the cache's way, and pick a label. See the FAQ for why "one request per label" is the right shape.