Rijksmuseum Mcp

Choosing between search_artwork(provenance=...) and search_provenance: search_artwork with provenance is a keyword search over raw provenance text — fast for counting, faceting, and cross-tabulation (e.g. "how many Mannheimer works are porcelain?"). search_provenance returns structured, parsed chains with dates, prices, transfer types, and ownership periods — use it when you need to reason about the sequence of ownership, filter by event type, or rank by price or duration. For the last link in the chain — how the Rijksmuseum acquired a work — also check creditLine, which covers ~358K artworks (vs ~48K with parsed provenance) and often names donors or funds absent from the provenance chain (e.g. "Drucker-Fraser", "Vereniging Rembrandt").

Critical Parameter Distinctions

productionPlace vs depictedPlace

These are semantically opposite and getting them wrong produces systematically wrong results.

• productionPlace: where the object was made — "Amsterdam", "Delft"
• depictedPlace: what place the object shows — "Batavia", "Nagasaki"

A Dutch painting of Batavia was made in Amsterdam. Use both together to distinguish objects from Asia versus European images of Asia.

subject vs iconclass

• subject: morphological stemming against ~108K terms (almost all in English); best first pass; try natural phrases ("winter landscape", "vanitas", "civic guard")
• iconclass: the Iconclass classification system — ~40,675 notations across 13 languages. A retrieval tool, not a descriptive language: a notation's meaning comes from its position in the hierarchy, not just its label. Use the Iconclass server (search, browse, resolve, expand_keys, search_prefix) to discover notation codes by keyword, concept, or hierarchy navigation — see its SKILL file for full workflows and query patterns. Each result includes collection counts signalling how many artworks carry that notation. Pass the code to search_artwork's iconclass parameter for precise filtering.

Decision rule: start with subject — it's faster and handles most queries well. Switch to iconclass when:

• you need to distinguish closely related scenes (Crucifixion vs Deposition, Annunciation vs Visitation) — these are sibling notations with distinct codes
• subject returns too broad a result set and you need hierarchical precision
• you want to explore a conceptual neighbourhood (browse children/siblings of a notation)
• the query is in a non-English language (Iconclass covers 13 languages; subject vocab is almost entirely English)
• you want to search by meaning rather than keywords (semantic mode: "religious suffering", "festive gathering")
• you need to combine multiple iconographic concepts in a single query (AND-combined codes)

dateMatch — controlling how date ranges are binned

Most artworks have date ranges (e.g. "1630–1640"), not exact years. dateMatch controls how these ranges interact with creationDate wildcards and collection_stats decade/century dimensions:

Rule of thumb: use midpoint whenever you're comparing counts across bins (decade trends, century breakdowns, gender ratios by period). Use the default overlaps for discovery queries where inclusiveness matters more than precision.

Available on search_artwork, semantic_search, and implicitly via collection_stats creationDateFrom/creationDateTo filters.

attributionQualifier + creator — structural limitation

The parameter description suggests combining attributionQualifier with creator to filter by source artist (e.g. "find all followers of Rembrandt"). This combination never works — the tool schema documentation is incorrect. Verified across all qualifier types (follower of, workshop of, circle of, attributed to): the structured creator field is always Unknown [painter] or anonymous, regardless of qualifier. The source artist's name appears only in the composite display string and is not exposed as a searchable entity field.

The correct strategy by qualifier type:

Additional complication — "manner of" is not a valid qualifier value. The Rijksmuseum uses "manner of" as a distinct attribution category, but "manner of" is not in the attributionQualifier controlled vocabulary and returns zero results. Works attributed in this way are retrievable only via aboutActor.

Canonical name form matters. The Rijksmuseum catalogue uses historical Dutch/Latin spellings for some artists. Bosch is catalogued as "Jheronimus Bosch", not "Hieronymus Bosch". Always check get_artwork_details on a known work to confirm the canonical form before filtering.

creator vs aboutActor vs depictedPerson

• creator: who made it — uses canonical name forms ("Rembrandt van Rijn")
• depictedPerson: who is shown in it — matches 210K name variants including historical forms
• aboutActor: broader cross-field search; more tolerant of name variants and cross-language forms ("Louis XIV" finds "Lodewijk XIV")

Title language coverage — the catalogue is mostly Dutch

The Rijksmuseum has translated only ~4% of artwork titles into English (~35K of 833K works). The remaining ~96% have Dutch titles only — typically medals, coins, prints, decorative arts, and other minor objects. Famous paintings and curatorially-prominent works almost always have both languages; the long tail does not.

Implications for research:

• Result presentation: when reporting results to the user, use whatever title language exists rather than insisting on English. For an 18th-century medal, "Wilhelmina Koningin der Nederlanden" is the canonical title — there is no English version, and inventing one would misrepresent the catalogue.
• Ranking: title BM25 scores are computed across both languages, so an English title query against a Dutch-only object scores zero on title (but the object may still surface via description, subject, iconclass, or semantic_search, all of which have stronger English coverage).

search_provenance: two query layers

search_provenance has two data layers that answer fundamentally different questions.

• layer: "events" (default): individual transactions — each with a date, location, price, parties (with roles and positions), and transfer type. Think of it as a ledger of what happened. Events-only parameters: transferType, excludeTransferType, hasPrice, currency, hasGap, relatedTo.
• layer: "periods": interpreted ownership spans — who held the work, how they acquired it, and for how long. Think of it as a timeline of who owned what. Periods-only parameters: ownerName, acquisitionMethod, minDuration, maxDuration, sortBy: "duration".

Shared parameters work on both layers: party, location, creator, dateFrom/dateTo, objectNumber, categoryMethod, positionMethod, sortBy, offset, facets.

dateFrom/dateTo semantics differ by layer:

• Events: filters on the event's date_year — "something happened between these years"
• Periods: dateFrom filters on begin_year, dateTo on end_year — "ownership that started after X AND ended before Y"

The periods interpretation is much more restrictive — dateFrom=1933, dateTo=1945 on periods misses any ownership that started before 1933 or extended past 1945. For date-range queries (especially wartime provenance), prefer the events layer.

Anti-join pattern (transferType + excludeTransferType): artwork-level set difference. transferType: "confiscation", excludeTransferType: "restitution" returns artworks that were confiscated but never restituted. Note: items recuperated (recovered by Allied forces) are not the same as items restituted (formally returned to original owners) — they will appear in anti-join results.

Filter requirement: both layers reject bare queries. At least one content filter is required. If you need a collection-wide ranking, use a broad filter such as dateFrom: 1400 as a catch-all.

For the full provenance data model — AAM text format, transfer type vocabulary, party roles and positions, date/currency representations, and tested query patterns — see references/provenance-patterns.md.

Full-text filters vs vocabulary filters (ranking matters)

Results rank by BM25 text relevance when any of these are active: title, description, inscription, provenance, creditLine, curatorialNarrative.

Results rank by importance (image availability + curatorial attention + metadata richness) when only vocabulary or column filters are active.

Practical rule: if you want the most significant works to surface first, use vocabulary filters alone. If you want keyword relevance, activate a text filter.

Modifier Parameters (cannot stand alone)

These narrow results but require at least one other content filter:

Key Workflows

1. Scope Before You Browse

For any counting, distributional, or comparative question, start with collection_stats — it answers in one call what would otherwise require multiple compact: true loops.

# Example: Rembrandt's output across media ��� one call
collection_stats(dimension="type", creator="Rembrandt van Rijn")
# → painting 314 (38.2%), print 289 (35.2%), drawing 218 (26.5%)

# Cross-domain: what types of artworks have provenance events in Amsterdam?
collection_stats(dimension="type", hasProvenance=true, location="Amsterdam")

# Top 30 depicted persons
collection_stats(dimension="depictedPerson", topN=30)

# Cross-filter with Iconclass or curated sets
collection_stats(dimension="type", iconclass="73D82")
# → what types of artworks depict the Crucifixion?

collection_stats(dimension="creator", collectionSet="Japanese prints", topN=20)
# → top 20 creators in the Japanese prints set

# Physical size distributions (binned by binWidth, in cm)
collection_stats(dimension="height", type="painting", binWidth=20)
# → height distribution of paintings in 20cm bins

Use compact: true on search_artwork only when you need the actual object numbers (e.g. to feed into get_artwork_details), not just the count or distribution. For pure counting, collection_stats is always more efficient.

2. Iconclass Research

Never pass iconographic concepts as free text to search_artwork(iconclass=...) — it expects exact notation codes. Use the Iconclass server to discover codes first (see its SKILL file for full discovery workflows, FTS query patterns, hierarchy navigation, and cross-branch strategies).

Quick reference — Iconclass server tools:

• search(query=...) — keyword lookup ("crucifixion" → 73D6)
• search(semanticQuery=...) — concept search ("domestic animals" → 34B1)
• browse(notation=...) — hierarchy navigation (children, cross-refs, path)
• resolve(notation=[...]) — batch lookup of known codes
• expand_keys / search_prefix — key variants and subtree enumeration

Before handing off, check collectionCounts — a code with 0 artworks returns nothing; a code with 2,000 vs 3 signals very different curatorial depth. Never truncate Iconclass discovery queries — use the default maxResults (25) or higher so you can evaluate all returned notations and their counts before deciding which codes to hand off.

Searching with codes on this server:

search_artwork(iconclass=["73D82"])
search_artwork(iconclass=["73D82", "25F33(DOVE)"])  # AND-combines across branches

Multiple codes AND-combine — this is how you express compound iconographic queries. If results are unexpected, use the Iconclass server's semantic search to discover alternative branches — the same concept can live in multiple places (a dog as pet under 34B11, a dog as symbol of fidelity under 11A(DOG)).

3. Century / Decade Wildcards

For longitudinal analysis, prefer collection_stats with decade or century binning. Use dateMatch: "midpoint" for statistical comparisons — the default overlaps mode double-counts broadly-dated objects across bins (see Critical Parameter Distinctions above):

collection_stats(dimension="decade", technique="etching", creationDateFrom=1500, creationDateTo=1800)
# → decade-by-decade etching counts in one call (default overlaps — inclusive but inflated)

collection_stats(dimension="decade", technique="etching", creationDateFrom=1500, creationDateTo=1800, binWidth=50)
# → half-century bins

For queries that need individual object numbers (not just counts), use creationDate wildcards with search_artwork:

search_artwork(technique="etching", creationDate="16*", dateMatch="midpoint")
search_artwork(technique="etching", creationDate="17*", dateMatch="midpoint")
# dateMatch="midpoint" ensures each artwork is counted in exactly one century

Supported patterns: "1642" (exact year), "164*" (decade), "16*" (century), "1*" (millennium).

4. Semantic Search + Structured Verification

Use semantic_search for concepts with no vocabulary term or Iconclass code — atmosphere, emotion, cultural interpretation. Always follow up with a structured search_artwork to test whether the same works are reachable through controlled vocabulary.

semantic_search(query="loneliness and isolation in a vast empty space", type="painting")
# Review source text for each result — ask why it appeared
# Then: search_artwork(subject="landscape", curatorialNarrative="solitary")
# Compare overlap: works only reachable via semantic search are the interesting gap cases

Pre-filtering: semantic_search accepts all vocabulary filters from search_artwork as pre-filters — subject, iconclass, depictedPerson, depictedPlace, productionPlace, collectionSet, aboutActor, imageAvailable, dateMatch. These narrow the candidate set before semantic ranking, combining structured precision with conceptual search:

semantic_search(query="vanitas symbolism", subject="skull", type="painting")
# → paintings with skull subjects, ranked by how well they match "vanitas symbolism"

Filter specificity matters. Very broad single filters (e.g. type: "print" alone — 421K works, or material: "paper" — 643K) exceed the internal candidate limit, so semantic ranking operates on a subset. Results are still high-quality (distances are near-optimal), but a different subset might surface equally-good alternatives. For best coverage, combine two or more filters or pair a broad filter with a specific one (e.g. type: "print", subject: "landscape").

Language note: English queries yield slightly higher precision against the bilingual catalogue even though the embedding model is multilingual. If a Dutch or German query returns unexpected results, reformulate in English.

5. Image Inspection and Overlay Placement

Basic inspection + zoom is now a single step. inspect_artwork_image automatically navigates the open viewer to the inspected region (navigateViewer defaults to true). No separate navigate_viewer call is needed for basic zoom — the viewer stays in sync with your analysis.

# Single call: inspect a region AND zoom the viewer there
inspect_artwork_image(objectNumber="SK-C-5", region="pct:70,60,20,20")
# → base64 image for AI analysis + viewer auto-zooms to the same region

For overlays, use the inspect → overlay two-step with relativeTo:

# Step 1: inspect the area
inspect_artwork_image(objectNumber="SK-C-5", region="pct:70,60,20,20")
# → found a detail at roughly the center of this crop

# Step 2: place overlay using crop-local coordinates — the server projects to full-image space
navigate_viewer(viewUUID=..., commands=[{
  action: "add_overlay",
  region: "pct:30,30,40,40",      # coordinates within the crop
  relativeTo: "pct:70,60,20,20",  # the crop region from Step 1
  label: "Detail"
}])

The relativeTo parameter eliminates manual coordinate conversion — specify where the feature is within the crop, and the server handles the projection. Both region and relativeTo must use pct: format.

For pixel-precision work on dense compositions (e.g. cataloguing several insects on a flower piece), use crop_pixels:x,y,w,h instead — coordinates are full-image pixels, so no relativeTo projection is needed. inspect_artwork_image echoes nativeWidth/nativeHeight in both the structured payload and the caption so you can bound pixel values to the actual image.

Use region="full" for an initial composition overview before cropping to details. inspect_artwork_image can surface content absent from structured metadata — unsigned Japanese prints often have readable artist signatures, publisher seals, and poem cartouches that the catalogue has not transcribed.

6. Provenance and Acquisition Research

Provenance research moves through three levels of detail:

1. Scope and profile — search_artwork with provenance (keyword) or creditLine for fast counts and facets. creditLine covers ~358K artworks (vs ~48K with parsed provenance) and captures the last link: how the museum acquired the work. Combine with productionPlace + expandPlaceHierarchy for geographic cross-tabulation.

2. Structured chain analysis — search_provenance for parsed chains with dates, prices, transfer types, and ownership periods. Add facets: true for quick distributional context alongside chain results.

3. Single-object deep dive — search_provenance(objectNumber=...) for the full chain, then get_artwork_details for narrative text and curatorial context.

Cross-domain queries: hasProvenance: true on search_artwork or collection_stats bridges the two systems — e.g. collection_stats(dimension="type", hasProvenance=true).

Enrichment transparency: categoryMethod and positionMethod are queryable input filters on search_provenance, not just output fields. When results contain LLM-enriched records, the response includes a review URL. Always show this URL to the user.

Pagination: when totalResults exceeds 50, paginate with offset (increments of 50 until offset ≥ totalResults).

For the complete data model (AAM text format, transfer type vocabulary, party roles, date and currency representations), tested query patterns (collector profiling, wartime provenance, price history, acquisition channels, decade-level time series), and enrichment methodology, see references/provenance-patterns.md.

7. Source–Copy Navigation

To trace reproductive prints back to their painted sources, use productionRole to find copies, then search for the source artist's original:

search_artwork(productionRole="after painting by", creator="Rembrandt van Rijn")
# → get_artwork_details on a result to read its description (often names the source)
# → search_artwork(creator="Rembrandt van Rijn", type="painting", title="...") to find the source
# → get_artwork_image on both for side-by-side comparison

Note: relatedObjects currently returns an empty array (not available from the vocabulary database). Use description text and title matching to navigate between related works.

8. Collection Depth Assessment

For grant applications or scoping a research site:

collection_stats(dimension="creator", type="print", productionPlace="Japan", topN=20)
# → top 20 print artists from Japan + total count, in one call

collection_stats(dimension="decade", type="print", productionPlace="Japan")
# → temporal distribution

list_curated_sets(query="Japan")                                       # curatorial groupings
browse_set(setSpec="...")                                               # range of artists/dates
search_artwork(productionPlace="Japan", type="print", maxResults=10)  # sample works for closer inspection

9. Gender and Demographic Analysis

creatorGender is a modifier — always combine it with at least one content filter. Use century wildcards to reveal structural patterns over time.

# Count women painters per century — use dateMatch="midpoint" to avoid double-counting
search_artwork(creatorGender="female", type="painting", creationDate="17*", dateMatch="midpoint", compact=true)
# Repeat for "18*", "19*"; then drop compact=true to identify leading artists per century

# Or: one-call gender breakdown per century
collection_stats(dimension="creatorGender", type="painting", creationDateFrom=1700, creationDateTo=1799)
# → male 4,521 (96.3%), female 48 (1.0%), ...

Known Limitations and Fallbacks

Output Conventions

• Always report objectNumber (e.g. SK-C-5) linked to the Rijksmuseum's page for that artwork when surfacing works — it is the stable identifier across all tools
• Use persistentId (handle.net URI) for any citation or bibliographic reference
• When presenting multiple works, lead with the most significant first (the default importance ranking handles this for vocabulary queries)
• For image inspection findings, distinguish explicitly between what the structured metadata says and what the AI reads directly from the image — the distinction matters for research rigour