@lde/search-typesense 0.4.3 → 0.5.0

package/README.md

@@ -26,20 +26,28 @@ they cost no RAM; only the folded `*_search_${locale}`, facet/reference and
 `*_sort_${locale}` companions are indexed. Keeping retrieval-only fields
 un-indexed is the lever for holding a large index’s RAM down.
 
-`createTypesenseSearchEngine(client, { collection, labelsCollection })` is the
+`createTypesenseSearchEngine(client, schema, { collections })` is the
 `SearchEngine` implementation. Each search:
 
 - validates the query against the search type (the port contract — a
   structurally invalid query is rejected, never sent);
 - compiles it into Typesense search params (`buildSearchParams`);
 - runs the search;
-- resolves reference (and reference-facet) labels from the sidecar `labels`
-  collection in a single lookup;
+- resolves reference (and reference-facet) labels **per reference field**
+  from the collection of the `SearchType` its `labelSource` names – all
+  sources bundled into a single lookup. A reference without a `labelSource`
+  stays id-only. With `labelCacheTtlMs` set, each label-source collection is
+  instead loaded once into an in-memory cache;
 - reconstructs the logical `SearchResult` (`parseSearchResponse`) — language
   maps, labelled references, labelled facet buckets.
 
+A label source is just another `SearchType` in the schema (with an `output`,
+`searchable` text field called `label`) whose collection appears in
+`collections` – a typed entity collection and a ‘labels collection’ are the
+same kind of thing.
+
 `searchFacets` – the port’s batch entry point – answers a whole batch of
-facet-only queries (e.g. a faceted sidebar’s skip-own-filter variants) as a
+facet-only queries (e.g. a faceted listing’s skip-own-filter variants) as a
 **single `multi_search` round-trip**, with one bundled label lookup shared by
 every facet result in the batch. A failed entry is reported in place as a
 per-query outcome, so its siblings’ facets survive.

package/dist/search.d.ts

@@ -1,42 +1,53 @@
 import type { Client } from 'typesense';
-import { type LocalizedValue, type SearchEngine, type SearchResult, type SearchSchema, type SearchType } from '@lde/search';
+import { type LocalizedValue, type SearchEngine, type SearchResult, type SearchSchema, type SearchType, type TextField } from '@lde/search';
 import { type BuildSearchParamsOptions } from './query-compiler.js';
-/** Where the engine reads documents and (optionally) reference labels — plus
- *  every query-compiler knob ({@link BuildSearchParamsOptions}), declared once
- *  there and forwarded wholesale into each search. `TypeName` is the union of
- *  the schema’s type names, so a literal schema makes the `collections` map
- *  exhaustive at compile time. */
+/** Where the engine reads documents — plus every query-compiler knob
+ *  ({@link BuildSearchParamsOptions}), declared once there and forwarded
+ *  wholesale into each search. `TypeName` is the union of the schema’s type
+ *  names, so a literal schema makes the `collections` map exhaustive at
+ *  compile time. Reference labels resolve per field from the collection of
+ *  the SearchType its `labelSource` names — a label source is just another
+ *  entry in `collections`. */
 export interface TypesenseSearchEngineOptions<TypeName extends string = string> extends BuildSearchParamsOptions {
     /** The collection (or alias) per search type, keyed by the type’s `name` —
      *  with a literal schema, omitting a type is a compile error. */
     readonly collections: Readonly<Record<TypeName, string>>;
-    /** The sidecar `labels` collection (IRI → label); omit for id-only references. */
-    readonly labelsCollection?: string;
     /**
      * Called when reference-label resolution fails; the search then degrades to
      * id-only references rather than failing. Optional — omit to swallow silently.
      */
     readonly onLabelError?: (error: unknown) => void;
     /**
-     * Opt-in in-memory label cache. When set (and {@link labelsCollection} is
-     * set), the FULL sidecar `labels` collection is loaded once via the documents
-     * export endpoint and held in a process-lifetime cache for this many
-     * milliseconds; each `search` then resolves its reference labels by in-memory
-     * lookup instead of a per-search `multi_search` round-trip. Omit to keep the
-     * per-search {@link fetchLabels} behaviour unchanged.
+     * Opt-in in-memory label cache. When set, each label-source collection is
+     * loaded in FULL once via the documents export endpoint and held in a
+     * process-lifetime cache for this many milliseconds; each `search` then
+     * resolves its reference labels by in-memory lookup instead of a per-search
+     * `multi_search` round-trip. Omit to keep the per-search
+     * {@link fetchLabels} behaviour unchanged.
      */
     readonly labelCacheTtlMs?: number;
 }
+/**
+ * One reference field’s resolved label source: the collection it reads, the
+ * source type’s `label` declaration (for reconstructing localized labels) and
+ * the comma-joined physical search fields (one per locale) a label search
+ * queries.
+ */
+export interface LabelSource {
+    readonly collection: string;
+    readonly labelField: TextField;
+    readonly queryBy: string;
+}
 /**
  * A Typesense-backed {@link SearchEngine}, bound to the whole
  * {@link SearchSchema} at construction — like every other schema consumer.
- * Each type’s collection comes from `options.collections`; the label sidecar
- * and its cache are shared across all types. `search` compiles the query
- * ({@link buildSearchParams}), runs it against the type’s collection,
- * resolves the reference labels for the page of hits from the sidecar
- * `labels` collection in one lookup, and reconstructs the engine-neutral
+ * Each type’s collection comes from `options.collections`. `search` compiles
+ * the query ({@link buildSearchParams}), runs it against the type’s
+ * collection, resolves the reference labels for the page of hits — each
+ * reference field from its own label source’s collection, all sources bundled
+ * into one lookup — and reconstructs the engine-neutral
  * {@link SearchResult} ({@link parseSearchResponse}). `searchFacets` answers
- * a whole facet batch (e.g. a sidebar’s skip-own-filter query variants) as a
+ * a whole facet batch (e.g. a faceted listing’s skip-own-filter query variants) as a
  * single `multi_search` round-trip with one shared label lookup; a failed
  * entry is reported in place as a per-query outcome, so it never discards
  * its siblings’ facets. Every engine specific stays here; consumers see only
@@ -47,20 +58,26 @@ export interface TypesenseSearchEngineOptions<TypeName extends string = string>
  * facet/document keys per call — no caller-side generics.
  */
 export declare function createTypesenseSearchEngine<const Types extends readonly SearchType[]>(client: Client, schema: SearchSchema<Types>, options: TypesenseSearchEngineOptions<Types[number]['name']>): SearchEngine<Types>;
+/** One label lookup: a source and the distinct IRIs to resolve against it. */
+export interface LabelLookupGroup {
+    readonly source: LabelSource;
+    readonly iris: readonly string[];
+}
 /**
- * Resolve labels for `iris` from the sidecar `labels` collection. Each
- * `label_${locale}` becomes a language-map entry; the default `label` is the
- * untagged (`und`) fallback when no locale variant exists.
+ * Resolve labels from each group’s label-source collection. Localized labels
+ * are reconstructed from the source type’s `label` declaration; a bare
+ * untagged `label` value is the `und` fallback when no locale variant exists.
  *
- * Sent as one `multi_search` (POST) call, the id-list split over per-search
- * batches: the id-list of a page or facet carrying many references — e.g. a
- * dataset with dozens of classes — would overflow Typesense’s GET query-string
- * limit (4000 chars, and IRIs URL-encode to several times their length) if it
- * travelled in the URL. POST puts it in the body; each batch stays under
- * Typesense’s `per_page` cap, and bundling the batches keeps it one round-trip
- * regardless of IRI count. Exported for unit testing against a fake client.
+ * All groups travel as ONE `multi_search` (POST) call, each group’s id-list
+ * split over per-search batches: the id-list of a page or facet carrying many
+ * references — e.g. a dataset with dozens of classes — would overflow
+ * Typesense’s GET query-string limit (4000 chars, and IRIs URL-encode to
+ * several times their length) if it travelled in the URL. POST puts it in the
+ * body; each batch stays under Typesense’s `per_page` cap, and bundling the
+ * batches keeps it one round-trip regardless of IRI or source count. Exported
+ * for unit testing against a fake client.
  */
-export declare function fetchLabels(client: Pick<Client, 'multiSearch'>, collection: string, iris: readonly string[]): Promise<Map<string, LocalizedValue>>;
+export declare function fetchLabels(client: Pick<Client, 'multiSearch'>, groups: readonly LabelLookupGroup[], onError?: (error: Error) => void): Promise<Map<string, LocalizedValue>>;
 /** The subset of a Typesense search response this adapter reads. */
 export interface TypesenseSearchResponse {
     readonly found: number;
@@ -79,7 +96,7 @@ export interface TypesenseSearchResponse {
  * Reconstruct a Typesense response into the engine-neutral {@link SearchResult}:
  * the flat, fanned-out document is turned back into a logical one (per-locale
  * display fields → a language map, reference IRIs → labelled references via the
- * sidecar `labels` lookup, scalars passed through). `labels` maps a reference IRI
+ * label-source lookup, scalars passed through). `labels` maps a reference IRI
  * to its resolved label; an IRI absent from it yields an id-only reference.
  */
 export declare function parseSearchResponse(response: TypesenseSearchResponse, searchType: SearchType, labels: ReadonlyMap<string, LocalizedValue>): SearchResult;

package/dist/search.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"search.d.ts","sourceRoot":"","sources":["../src/search.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,WAAW,CAAC;AACxC,OAAO,EAGL,KAAK,cAAc,EAMnB,KAAK,YAAY,EAEjB,KAAK,YAAY,EACjB,KAAK,YAAY,EACjB,KAAK,UAAU,EAGhB,MAAM,aAAa,CAAC;AAUrB,OAAO,EAGL,KAAK,wBAAwB,EAC9B,MAAM,qBAAqB,CAAC;AAE7B;;;;kCAIkC;AAClC,MAAM,WAAW,4BAA4B,CAC3C,QAAQ,SAAS,MAAM,GAAG,MAAM,CAChC,SAAQ,wBAAwB;IAChC;qEACiE;IACjE,QAAQ,CAAC,WAAW,EAAE,QAAQ,CAAC,MAAM,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC,CAAC;IACzD,kFAAkF;IAClF,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,CAAC;IACnC;;;OAGG;IACH,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC,KAAK,EAAE,OAAO,KAAK,IAAI,CAAC;IACjD;;;;;;;OAOG;IACH,QAAQ,CAAC,eAAe,CAAC,EAAE,MAAM,CAAC;CACnC;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAgB,2BAA2B,CACzC,KAAK,CAAC,KAAK,SAAS,SAAS,UAAU,EAAE,EAEzC,MAAM,EAAE,MAAM,EACd,MAAM,EAAE,YAAY,CAAC,KAAK,CAAC,EAC3B,OAAO,EAAE,4BAA4B,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,MAAM,CAAC,CAAC,GAC3D,YAAY,CAAC,KAAK,CAAC,CAwKrB;AA+DD;;;;;;;;;;;;GAYG;AACH,wBAAsB,WAAW,CAC/B,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,aAAa,CAAC,EACnC,UAAU,EAAE,MAAM,EAClB,IAAI,EAAE,SAAS,MAAM,EAAE,GACtB,OAAO,CAAC,GAAG,CAAC,MAAM,EAAE,cAAc,CAAC,CAAC,CAmCtC;AA6BD,oEAAoE;AACpE,MAAM,WAAW,uBAAuB;IACtC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,IAAI,CAAC,EAAE,SAAS;QAAE,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAAE,EAAE,CAAC;IAC1E,QAAQ,CAAC,YAAY,CAAC,EAAE,SAAS;QAC/B,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;QAC5B,QAAQ,CAAC,MAAM,EAAE,SAAS;YACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;YACvB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;SACxB,EAAE,CAAC;KACL,EAAE,CAAC;CACL;AAED;;;;;;GAMG;AACH,wBAAgB,mBAAmB,CACjC,QAAQ,EAAE,uBAAuB,EACjC,UAAU,EAAE,UAAU,EACtB,MAAM,EAAE,WAAW,CAAC,MAAM,EAAE,cAAc,CAAC,GAC1C,YAAY,CAiCd"}
+{"version":3,"file":"search.d.ts","sourceRoot":"","sources":["../src/search.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,WAAW,CAAC;AACxC,OAAO,EAGL,KAAK,cAAc,EAMnB,KAAK,YAAY,EAEjB,KAAK,YAAY,EACjB,KAAK,YAAY,EACjB,KAAK,UAAU,EAEf,KAAK,SAAS,EACf,MAAM,aAAa,CAAC;AAWrB,OAAO,EAGL,KAAK,wBAAwB,EAC9B,MAAM,qBAAqB,CAAC;AAE7B;;;;;;8BAM8B;AAC9B,MAAM,WAAW,4BAA4B,CAC3C,QAAQ,SAAS,MAAM,GAAG,MAAM,CAChC,SAAQ,wBAAwB;IAChC;qEACiE;IACjE,QAAQ,CAAC,WAAW,EAAE,QAAQ,CAAC,MAAM,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC,CAAC;IACzD;;;OAGG;IACH,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC,KAAK,EAAE,OAAO,KAAK,IAAI,CAAC;IACjD;;;;;;;OAOG;IACH,QAAQ,CAAC,eAAe,CAAC,EAAE,MAAM,CAAC;CACnC;AAED;;;;;GAKG;AACH,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,UAAU,EAAE,SAAS,CAAC;IAC/B,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAED;;;;;;;;;;;;;;;;;;GAkBG;AACH,wBAAgB,2BAA2B,CACzC,KAAK,CAAC,KAAK,SAAS,SAAS,UAAU,EAAE,EAEzC,MAAM,EAAE,MAAM,EACd,MAAM,EAAE,YAAY,CAAC,KAAK,CAAC,EAC3B,OAAO,EAAE,4BAA4B,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,MAAM,CAAC,CAAC,GAC3D,YAAY,CAAC,KAAK,CAAC,CAsRrB;AAED,8EAA8E;AAC9E,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,MAAM,EAAE,WAAW,CAAC;IAC7B,QAAQ,CAAC,IAAI,EAAE,SAAS,MAAM,EAAE,CAAC;CAClC;AA0FD;;;;;;;;;;;;;GAaG;AACH,wBAAsB,WAAW,CAC/B,MAAM,EAAE,IAAI,CAAC,MAAM,EAAE,aAAa,CAAC,EACnC,MAAM,EAAE,SAAS,gBAAgB,EAAE,EACnC,OAAO,CAAC,EAAE,CAAC,KAAK,EAAE,KAAK,KAAK,IAAI,GAC/B,OAAO,CAAC,GAAG,CAAC,MAAM,EAAE,cAAc,CAAC,CAAC,CAgDtC;AAgDD,oEAAoE;AACpE,MAAM,WAAW,uBAAuB;IACtC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,IAAI,CAAC,EAAE,SAAS;QAAE,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAAE,EAAE,CAAC;IAC1E,QAAQ,CAAC,YAAY,CAAC,EAAE,SAAS;QAC/B,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;QAC5B,QAAQ,CAAC,MAAM,EAAE,SAAS;YACxB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;YACvB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;SACxB,EAAE,CAAC;KACL,EAAE,CAAC;CACL;AAED;;;;;;GAMG;AACH,wBAAgB,mBAAmB,CACjC,QAAQ,EAAE,uBAAuB,EACjC,UAAU,EAAE,UAAU,EACtB,MAAM,EAAE,WAAW,CAAC,MAAM,EAAE,cAAc,CAAC,GAC1C,YAAY,CAsCd"}

package/dist/search.js

@@ -1,15 +1,15 @@
-import { assertTypeInSchema, assertValidQuery, fieldNamed, isRangeFacet, outputFields, physicalFields, referenceFields, } from '@lde/search/adapter';
+import { assertTypeInSchema, assertValidQuery, fieldNamed, isRangeFacet, labelFieldOf, outputFields, physicalFields, referenceFields, } from '@lde/search/adapter';
 import { buildSearchParams, escapeFilterValue, } from './query-compiler.js';
 /**
  * A Typesense-backed {@link SearchEngine}, bound to the whole
  * {@link SearchSchema} at construction — like every other schema consumer.
- * Each type’s collection comes from `options.collections`; the label sidecar
- * and its cache are shared across all types. `search` compiles the query
- * ({@link buildSearchParams}), runs it against the type’s collection,
- * resolves the reference labels for the page of hits from the sidecar
- * `labels` collection in one lookup, and reconstructs the engine-neutral
+ * Each type’s collection comes from `options.collections`. `search` compiles
+ * the query ({@link buildSearchParams}), runs it against the type’s
+ * collection, resolves the reference labels for the page of hits — each
+ * reference field from its own label source’s collection, all sources bundled
+ * into one lookup — and reconstructs the engine-neutral
  * {@link SearchResult} ({@link parseSearchResponse}). `searchFacets` answers
- * a whole facet batch (e.g. a sidebar’s skip-own-filter query variants) as a
+ * a whole facet batch (e.g. a faceted listing’s skip-own-filter query variants) as a
  * single `multi_search` round-trip with one shared label lookup; a failed
  * entry is reported in place as a per-query outcome, so it never discards
  * its siblings’ facets. Every engine specific stays here; consumers see only
@@ -30,21 +30,69 @@ export function createTypesenseSearchEngine(client, schema, options) {
         }
         return [searchType.type, collection];
     }));
-    // Process-lifetime cache for the FULL `labels` collection, held in the engine
-    // closure. Populated lazily on the first cached search; `loadAll` is the
-    // single-flight in-flight promise so concurrent first-loads share one export.
-    let cachedLabels;
-    let cacheExpiresAt = 0;
-    let inFlightLoad;
-    function cachedAllLabels(labelsCollection, ttlMs) {
-        if (cachedLabels !== undefined && Date.now() < cacheExpiresAt) {
-            return Promise.resolve(cachedLabels);
+    // Resolve every reference field's label source ONCE at construction. The
+    // schema already guarantees the named type exists and serves labels
+    // (`searchSchema`/`labelFieldOf`), and `collections` is exhaustive per type.
+    const typesByName = new Map([...schema.values()].map((searchType) => [searchType.name, searchType]));
+    const labelSources = new Map([...schema.values()].map((searchType) => [
+        searchType.type,
+        new Map(referenceFields(searchType)
+            .filter((field) => field.labelSource !== undefined)
+            .map((field) => {
+            const source = typesByName.get(field.labelSource);
+            const labelField = labelFieldOf(source);
+            return [
+                field.name,
+                {
+                    collection: collections.get(source.type),
+                    labelField,
+                    queryBy: physicalFields(labelField).search.join(','),
+                },
+            ];
+        })),
+    ]));
+    // The distinct source collections per type, for the cached path – fixed at
+    // construction, so no per-search dedup dance.
+    const distinctLabelSources = new Map([...labelSources].map(([type, sources]) => [
+        type,
+        [
+            ...new Map([...sources.values()].map((source) => [source.collection, source])).values(),
+        ],
+    ]));
+    // The output reference fields that carry hit labels, each paired with its
+    // resolved label source – fixed at construction, so labelLookupGroups need
+    // not re-derive or re-resolve them on every search.
+    const outputReferenceSources = new Map([...schema.values()].map((searchType) => {
+        const sources = labelSources.get(searchType.type);
+        return [
+            searchType.type,
+            referenceFields(searchType)
+                .filter((field) => field.output === true && sources?.has(field.name))
+                .map((field) => ({
+                name: field.name,
+                source: sources.get(field.name),
+            })),
+        ];
+    }));
+    // Process-lifetime cache per label-source collection, held in the engine
+    // closure. Populated lazily on the first cached search; `inFlightLoads` is
+    // the single-flight promise per collection so concurrent first-loads share
+    // one export each.
+    const cachedLabels = new Map();
+    const inFlightLoads = new Map();
+    function cachedAllLabels(source, ttlMs) {
+        const cached = cachedLabels.get(source.collection);
+        if (cached !== undefined && Date.now() < cached.expiresAt) {
+            return Promise.resolve(cached.labels);
         }
         // Single-flight: a load already running serves every concurrent caller.
-        inFlightLoad ??= loadAllLabels(client, labelsCollection)
+        let load = inFlightLoads.get(source.collection);
+        load ??= loadAllLabels(client, source)
             .then((loaded) => {
-            cachedLabels = loaded;
-            cacheExpiresAt = Date.now() + ttlMs;
+            cachedLabels.set(source.collection, {
+                labels: loaded,
+                expiresAt: Date.now() + ttlMs,
+            });
             return loaded;
         })
             // A failed load degrades to id-only references and is NOT cached, so the
@@ -54,35 +102,52 @@ export function createTypesenseSearchEngine(client, schema, options) {
             return new Map();
         })
             .finally(() => {
-            inFlightLoad = undefined;
+            inFlightLoads.delete(source.collection);
         });
-        return inFlightLoad;
+        inFlightLoads.set(source.collection, load);
+        return load;
+    }
+    // The merged per-type view over the cached per-collection maps, reused
+    // until any constituent map reloads (same instances = still valid), so a
+    // multi-source type does not pay an O(all labels) merge per search.
+    const mergedCache = new Map();
+    function mergeCachedLabels(typeIri, parts) {
+        const cached = mergedCache.get(typeIri);
+        if (cached !== undefined &&
+            cached.parts.length === parts.length &&
+            cached.parts.every((part, index) => part === parts[index])) {
+            return cached.merged;
+        }
+        const merged = mergeLabels(parts);
+        mergedCache.set(typeIri, { parts, merged });
+        return merged;
     }
-    // Cached path: the once-loaded full collection serves labels by in-memory
+    // Cached path: the once-loaded full collections serve labels by in-memory
     // lookup (no per-search round-trip). The load does not depend on the
     // response, so it is started BEFORE awaiting the search and runs alongside
     // it; it never rejects (a failed load degrades to an empty map), so it
     // cannot leave an unhandled rejection behind if the search itself fails.
-    // `undefined` when the cache (or the sidecar) is not configured.
-    function startCachedLabels() {
-        return options.labelsCollection !== undefined &&
-            options.labelCacheTtlMs !== undefined
-            ? cachedAllLabels(options.labelsCollection, options.labelCacheTtlMs)
-            : undefined;
+    // `undefined` when the cache is off or the type has no label sources.
+    function startCachedLabels(searchType) {
+        const sources = distinctLabelSources.get(searchType.type);
+        if (options.labelCacheTtlMs === undefined ||
+            sources === undefined ||
+            sources.length === 0) {
+            return undefined;
+        }
+        const ttlMs = options.labelCacheTtlMs;
+        return Promise.all(sources.map((source) => cachedAllLabels(source, ttlMs))).then((maps) => mergeCachedLabels(searchType.type, maps));
     }
-    // Labels are supplementary: a failed lookup (e.g. the sidecar collection
+    // Labels are supplementary: a failed lookup (e.g. a label-source collection
     // mid-rebuild) degrades to id-only references rather than failing the whole
-    // search, so the listing still renders with bare IRIs. `iris` is a thunk so
-    // the cached and no-sidecar paths never pay for collecting them.
-    async function resolveLabels(cachedLabelsPromise, iris) {
+    // search, so the listing still renders with bare IRIs. `groups` is a thunk
+    // so the cached and source-less paths never pay for collecting the IRIs.
+    async function resolveLabels(cachedLabelsPromise, groups) {
         if (cachedLabelsPromise !== undefined) {
             return cachedLabelsPromise;
         }
-        if (options.labelsCollection === undefined) {
-            return new Map();
-        }
         try {
-            return await fetchLabels(client, options.labelsCollection, iris());
+            return await fetchLabels(client, groups(), options.onLabelError);
         }
         catch (error) {
             options.onLabelError?.(error);
@@ -98,12 +163,12 @@ export function createTypesenseSearchEngine(client, schema, options) {
             assertTypeInSchema(schema, searchType);
             assertValidQuery(query, searchType);
             const params = buildSearchParams(query, searchType, options);
-            const cachedLabelsPromise = startCachedLabels();
+            const cachedLabelsPromise = startCachedLabels(searchType);
             const response = (await client
                 .collections(collections.get(searchType.type))
                 .documents()
                 .search(params));
-            const labels = await resolveLabels(cachedLabelsPromise, () => referenceIris(response, searchType));
+            const labels = await resolveLabels(cachedLabelsPromise, () => labelLookupGroups([response], labelSources.get(searchType.type), outputReferenceSources.get(searchType.type) ?? []));
             return parseSearchResponse(response, searchType, labels);
         },
         async searchFacets(searchType, queries) {
@@ -115,7 +180,7 @@ export function createTypesenseSearchEngine(client, schema, options) {
                 return [];
             }
             const collection = collections.get(searchType.type);
-            const cachedLabelsPromise = startCachedLabels();
+            const cachedLabelsPromise = startCachedLabels(searchType);
             // The whole batch travels as ONE multi_search round-trip. Each query
             // compiles as facet-only regardless of what it carries: no hits
             // (per_page 0) and no ordering – nothing is transferred or sorted that
@@ -128,9 +193,7 @@ export function createTypesenseSearchEngine(client, schema, options) {
             }));
             // One label lookup serves every successful facet result in the batch.
             const responses = results.filter((result) => !('error' in result));
-            const labels = await resolveLabels(cachedLabelsPromise, () => [
-                ...new Set(responses.flatMap((response) => referenceIris(response, searchType))),
-            ]);
+            const labels = await resolveLabels(cachedLabelsPromise, () => labelLookupGroups(responses, labelSources.get(searchType.type), outputReferenceSources.get(searchType.type) ?? []));
             // multi_search reports a failed entry inline instead of rejecting the
             // call; pass that through as a per-query outcome – naming the facets,
             // not the position, since the batch order is the caller's internal –
@@ -146,121 +209,172 @@ export function createTypesenseSearchEngine(client, schema, options) {
     return engine;
 }
 /**
- * Load the FULL `labels` collection into a label map via the documents export
- * endpoint, which streams every document as JSONL (one JSON object per line).
- * Each line is reconstructed by {@link labelToLocalizedValue}, exactly as the
+ * Load a FULL label-source collection into a label map via the documents
+ * export endpoint, which streams every document as JSONL (one JSON object per
+ * line). Each line is reconstructed by {@link labelValue}, exactly as the
  * per-search {@link fetchLabels} path does for its `multi_search` hits.
  */
-async function loadAllLabels(client, collection) {
-    const jsonl = await client.collections(collection).documents().export();
+async function loadAllLabels(client, source) {
+    const jsonl = await client
+        .collections(source.collection)
+        .documents()
+        .export();
     const labels = new Map();
     for (const line of jsonl.split('\n')) {
         if (line.length === 0) {
             continue;
         }
         const document = JSON.parse(line);
-        labels.set(String(document.id), labelToLocalizedValue(document));
+        const label = labelValue(document, source.labelField);
+        if (label !== undefined) {
+            labels.set(String(document.id), label);
+        }
     }
     return labels;
 }
-/** Every distinct reference IRI whose label the result will actually use. */
-function referenceIris(response, searchType) {
-    const referenceFieldSet = new Set(referenceFields(searchType).map((field) => field.name));
-    // Hits only carry labels for OUTPUT reference fields: reconstructDocument skips
-    // non-output fields, so resolving a non-output reference's hit labels (e.g. a
-    // facet-only `class` with dozens of IRIs per hit) is pure waste.
-    const outputReferenceFields = referenceFields(searchType)
-        .filter((field) => field.output === true)
-        .map((field) => field.name);
-    const iris = new Set();
-    for (const hit of response.hits ?? []) {
-        for (const name of outputReferenceFields) {
-            const raw = hit.document[name];
-            if (Array.isArray(raw)) {
-                for (const value of raw) {
-                    iris.add(String(value));
+/**
+ * Group every reference IRI the result will actually use by its field’s label
+ * source, one group per source collection. Fields without a `labelSource`
+ * stay id-only, so their IRIs never travel.
+ */
+function labelLookupGroups(responses, sources, outputSources) {
+    if (sources === undefined || sources.size === 0) {
+        return [];
+    }
+    const irisByCollection = new Map();
+    const add = (source, iri) => {
+        let group = irisByCollection.get(source.collection);
+        if (group === undefined) {
+            group = { source, iris: new Set() };
+            irisByCollection.set(source.collection, group);
+        }
+        group.iris.add(iri);
+    };
+    for (const response of responses) {
+        // Hits only carry labels for OUTPUT reference fields (reconstructDocument
+        // skips non-output fields); `outputSources` pairs each with its resolved
+        // source, precomputed per type.
+        for (const hit of response.hits ?? []) {
+            for (const { name, source } of outputSources) {
+                const raw = hit.document[name];
+                if (Array.isArray(raw)) {
+                    for (const value of raw) {
+                        add(source, String(value));
+                    }
+                }
+                else if (typeof raw === 'string') {
+                    add(source, raw);
                 }
-            }
-            else if (typeof raw === 'string') {
-                iris.add(raw);
             }
         }
-    }
-    // Reference-facet bucket values are IRIs too (incl. facet-only references like
-    // `class`); resolve them in the same lookup.
-    for (const facet of response.facet_counts ?? []) {
-        if (referenceFieldSet.has(facet.field_name)) {
+        // Reference-facet bucket values are IRIs too (incl. facet-only references
+        // like `class`); resolve them in the same lookup. Skip a non-source facet
+        // (e.g. a keyword facet) in one check instead of probing every bucket.
+        for (const facet of response.facet_counts ?? []) {
+            const source = sources.get(facet.field_name);
+            if (source === undefined) {
+                continue;
+            }
             for (const bucket of facet.counts) {
-                iris.add(bucket.value);
+                add(source, bucket.value);
             }
         }
     }
-    return [...iris];
+    return [...irisByCollection.values()].map(({ source, iris }) => ({
+        source,
+        iris: [...iris],
+    }));
 }
 /**
- * Resolve labels for `iris` from the sidecar `labels` collection. Each
- * `label_${locale}` becomes a language-map entry; the default `label` is the
- * untagged (`und`) fallback when no locale variant exists.
+ * Resolve labels from each group’s label-source collection. Localized labels
+ * are reconstructed from the source type’s `label` declaration; a bare
+ * untagged `label` value is the `und` fallback when no locale variant exists.
  *
- * Sent as one `multi_search` (POST) call, the id-list split over per-search
- * batches: the id-list of a page or facet carrying many references — e.g. a
- * dataset with dozens of classes — would overflow Typesense’s GET query-string
- * limit (4000 chars, and IRIs URL-encode to several times their length) if it
- * travelled in the URL. POST puts it in the body; each batch stays under
- * Typesense’s `per_page` cap, and bundling the batches keeps it one round-trip
- * regardless of IRI count. Exported for unit testing against a fake client.
+ * All groups travel as ONE `multi_search` (POST) call, each group’s id-list
+ * split over per-search batches: the id-list of a page or facet carrying many
+ * references — e.g. a dataset with dozens of classes — would overflow
+ * Typesense’s GET query-string limit (4000 chars, and IRIs URL-encode to
+ * several times their length) if it travelled in the URL. POST puts it in the
+ * body; each batch stays under Typesense’s `per_page` cap, and bundling the
+ * batches keeps it one round-trip regardless of IRI or source count. Exported
+ * for unit testing against a fake client.
  */
-export async function fetchLabels(client, collection, iris) {
+export async function fetchLabels(client, groups, onError) {
     const labels = new Map();
-    if (iris.length === 0) {
-        return labels;
-    }
     const searches = [];
-    for (let start = 0; start < iris.length; start += LABEL_BATCH_SIZE) {
-        const batch = iris.slice(start, start + LABEL_BATCH_SIZE);
-        searches.push({
-            collection,
-            q: '*',
-            query_by: 'label',
-            filter_by: `id:[${batch.map(escapeFilterValue).join(',')}]`,
-            per_page: batch.length,
-        });
+    const groupPerSearch = [];
+    for (const group of groups) {
+        for (let start = 0; start < group.iris.length; start += LABEL_BATCH_SIZE) {
+            const batch = group.iris.slice(start, start + LABEL_BATCH_SIZE);
+            searches.push({
+                collection: group.source.collection,
+                q: '*',
+                query_by: group.source.queryBy,
+                filter_by: `id:[${batch.map(escapeFilterValue).join(',')}]`,
+                per_page: batch.length,
+            });
+            groupPerSearch.push(group);
+        }
+    }
+    if (searches.length === 0) {
+        return labels;
     }
     const { results } = (await client.multiSearch.perform({ searches }));
-    for (const result of results) {
-        // multi_search reports a failed entry inline instead of rejecting; throw
-        // so the caller's degradation path (onLabelError, id-only references)
-        // engages rather than mistaking the failure for absent labels.
+    results.forEach((result, index) => {
+        // multi_search reports a failed entry inline instead of rejecting. Isolate
+        // it: report the failed source and skip its entry so the other sources'
+        // labels still land, instead of blanking the whole page to id-only – its
+        // own references fall back to id-only as their IRIs stay absent.
         if ('error' in result) {
-            throw new Error(`Typesense label lookup failed${result.code !== undefined ? ` (${result.code})` : ''}: ${result.error}`);
+            const failure = new Error(`Typesense label lookup in “${groupPerSearch[index].source.collection}” failed${result.code !== undefined ? ` (${result.code})` : ''}: ${result.error}`);
+            onError?.(failure);
+            return;
         }
         for (const hit of result.hits ?? []) {
-            labels.set(String(hit.document.id), labelToLocalizedValue(hit.document));
+            const label = labelValue(hit.document, groupPerSearch[index].source.labelField);
+            if (label !== undefined) {
+                labels.set(String(hit.document.id), label);
+            }
         }
-    }
+    });
     return labels;
 }
 /** Typesense caps `per_page` at 250; the multi_search POST body holds the
  *  id-list comfortably, so resolve references in batches of this size. */
 const LABEL_BATCH_SIZE = 200;
-/** Turn a `labels` document into a language map (`label_${locale}` → locale). */
-function labelToLocalizedValue(document) {
-    const map = {};
-    for (const [key, value] of Object.entries(document)) {
-        if (key.startsWith('label_') && typeof value === 'string') {
-            map[key.slice('label_'.length)] = [value];
-        }
+/**
+ * Reconstruct a label document into a language map via the source type’s
+ * `label` declaration (its per-locale display fields), with a bare untagged
+ * `label` value as the `und` fallback when no locale variant exists, or
+ * `undefined` when the document carries no usable label – so the reference
+ * stays id-only rather than gaining an empty label.
+ */
+function labelValue(document, labelField) {
+    const localized = localizedValue(document, labelField);
+    if (localized !== undefined) {
+        return localized;
+    }
+    const untagged = document[labelField.name];
+    return typeof untagged === 'string' ? { und: [untagged] } : undefined;
+}
+/** Merge per-collection label maps into one IRI-keyed map (URI identity). */
+function mergeLabels(maps) {
+    if (maps.length === 1) {
+        return maps[0];
     }
-    if (Object.keys(map).length === 0 && typeof document.label === 'string') {
-        map.und = [document.label];
+    const merged = new Map();
+    for (const map of maps) {
+        for (const [iri, label] of map) {
+            merged.set(iri, label);
+        }
     }
-    return map;
+    return merged;
 }
 /**
  * Reconstruct a Typesense response into the engine-neutral {@link SearchResult}:
  * the flat, fanned-out document is turned back into a logical one (per-locale
  * display fields → a language map, reference IRIs → labelled references via the
- * sidecar `labels` lookup, scalars passed through). `labels` maps a reference IRI
+ * label-source lookup, scalars passed through). `labels` maps a reference IRI
  * to its resolved label; an IRI absent from it yields an id-only reference.
  */
 export function parseSearchResponse(response, searchType, labels) {
@@ -270,7 +384,12 @@ export function parseSearchResponse(response, searchType, labels) {
     }));
     // Reference facets are IRI-keyed; their buckets carry a resolved data label.
     // Plain facets (tokens, free strings) carry no label — the consumer owns display.
-    const referenceFacets = new Set(referenceFields(searchType).map((field) => field.name));
+    // Only reference facets with a label source get bucket labels; an id-only
+    // reference facet stays id-only even when a cached full-collection map holds
+    // its IRIs.
+    const referenceFacets = new Set(referenceFields(searchType)
+        .filter((field) => field.labelSource !== undefined)
+        .map((field) => field.name));
     const facets = {};
     for (const facet of response.facet_counts ?? []) {
         const labelled = referenceFacets.has(facet.field_name);
@@ -348,7 +467,10 @@ function referenceValue(flat, field, labels) {
     }
     const iris = Array.isArray(raw) ? raw : [String(raw)];
     const references = iris.map((iri) => {
-        const label = labels.get(iri);
+        // A reference without a label source is id-only by declaration; never
+        // attach a label, even if the (cached, full-collection) map happens to
+        // hold this IRI from another source.
+        const label = field.labelSource === undefined ? undefined : labels.get(iri);
         return label === undefined ? { id: iri } : { id: iri, label };
     });
     return field.array === true ? references : references[0];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lde/search-typesense",
-  "version": "0.4.3",
+  "version": "0.5.0",
   "description": "Typesense implementation of the @lde/search SearchEngine port: collection-schema builder, query compiler, label-resolving result reconstruction, and blue/green index lifecycle. Engine-specific (Typesense) but domain-agnostic.",
   "repository": {
     "url": "git+https://github.com/ldelements/lde.git",
@@ -25,7 +25,7 @@
     "!**/*.tsbuildinfo"
   ],
   "dependencies": {
-    "@lde/search": "^0.3.1",
+    "@lde/search": "^0.4.0",
     "@lde/text-normalization": "^0.1.1",
     "tslib": "^2.3.0",
     "typesense": "^3.0.6"
@@ -35,6 +35,6 @@
   },
   "peerDependencies": {
     "@lde/dataset": "^0.7.8",
-    "@lde/pipeline": "^0.33.3"
+    "@lde/pipeline": "^0.33.4"
   }
 }