Semantic Catalog: Overview & Applications

• Semantic Catalogs are metadata-rich systems that attach explicit semantics to resources, enhancing discovery, queryability, and interoperability.
• They leverage diverse formalisms, including RDF, OWL, and typed object catalogs, to automate catalog construction and ensure data consistency.
• They support advanced query execution with deterministic access controls and governance measures to maintain trust, quality, and sustainability.

A semantic catalog is a metadata-rich, machine-actionable catalog layer that represents not only the existence of resources but also their meaning, provenance, structure, relationships, constraints, and admissible uses. Across product search, enterprise analytics, observability, astronomy, dataspaces, and AI governance, the term denotes a family of systems that make heterogeneous resources discoverable, comparable, queryable, and maintainable by attaching explicit semantics to catalog objects rather than treating them as opaque records or raw schema names (Nguyen et al., 2011, Agrawal et al., 23 Jun 2026, McCormick et al., 2024, Arnold et al., 24 Jan 2025, Golpayegani et al., 2024, Cecconi et al., 17 Apr 2025).

1. Conceptual scope and defining properties

The literature does not present a single canonical definition of a semantic catalog; instead, it converges on several closely related formulations. In semantic interoperability work, a catalogue of semantic artefacts is defined as “a dedicated web-based system that fosters the availability, discoverability and long-term preservation and maintenance of semantic artefacts,” where a semantic artefact is “a machine-actionable formalisation (represented using appropriate formats and serialisations, including RDF and non-RDF standards) of a conceptualisation, enabling sharing and reuse by humans and machines” (Corcho et al., 2023). In product search, the catalog is framed as a semantic integration layer that reconciles merchant schemas, fuses redundant evidence, and continually grows the structured catalog from marketplace data (Nguyen et al., 2011). In enterprise text-to-SQL, the central claim is that “the right unit to retrieve is the semantic catalog, not the warehouse rows or even the raw schema alone” (Agrawal et al., 23 Jun 2026).

A recurring theme is that a semantic catalog organizes metadata under shared semantic categories rather than leaving it in loosely defined fields. One explicit formalization models a catalog mental model as a partition of the metadata-item set $C$ into disjoint subsets:

$$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$

This formulation underlies the 5W1H+R model, which partitions metadata into Who, What, When, Why, Where, How, and relationships (+R), and was reported to be more comprehensible and easier to use than derived mental models from Google Cloud Data Catalog and LinkedIn DataHub (Subramaniam et al., 2021).

The notion of “semantic” also varies in strength. Some systems are semantic in a formal Semantic Web sense, using RDF, OWL, SKOS, SHACL, or DCAT (Arnold et al., 24 Jan 2025, Golpayegani et al., 2024, Cecconi et al., 17 Apr 2025). Others are semantic in a practical operational sense because they are self-describing, typed, and queryable enough to support automated discovery and use; the Siril HEALpixel Catalog Format is described in exactly this way, as not semantic in the formal ontology sense but semantic in the practical sense that it encodes enough metadata to make the catalog self-describing and spatially searchable (Knagg-Baugh et al., 21 Jan 2025). This suggests that semantic catalogs are best understood as systems that make meaning operational, whether by formal semantics, typed metadata, or domain-specific structure.

2. Representational architectures and metadata models

Semantic catalogs span a wide range of internal representations. Felis represents astronomical catalog semantics and metadata as a rich Pydantic data model expressed in human-readable and editable YAML; all schema objects have common attributes name, identifier, and description, a schema must contain at least one table, a table must contain one or more columns, and columns require a datatype (McCormick et al., 2024). Felis also maintains semantic properties such as units and IVOA Unified Content Descriptors, supports primary key, foreign key, and unique constraints, and can generate DDL for MySQL and PostgreSQL or populate TAP_SCHEMA for IVOA TAP services (McCormick et al., 2024).

AICat adopts a linked-data cataloguing profile. It is a thin-layer application profile built on DCAT v3, introduces aicat:Catalog as a subclass of dcat:Catalog, and reuses airo:AISystem, airo:AIModel, and airo:Data as DCAT-compatible resource classes. It further introduces aicat:system, aicat:model, and aicat:dataset as sub-properties of dcat:resource, and maps AI Act registration fields onto reused vocabularies such as DCTERMS, AIRO, DPV, DPV TECH, ODRL, and AIUP (Golpayegani et al., 2024). XFSC uses a different semantic stack: participants publish Self-Descriptions as W3C Verifiable Presentations containing Verifiable Credentials in JSON-LD; claims are RDF triples, validated against ontologies, SHACL shapes, and SKOS vocabularies, then extracted into a Neo4j graph via neosemantics and queried with openCypher (Arnold et al., 24 Jan 2025).

Other systems favor typed object catalogs or hierarchical storage engines. Schema-First Retrieval indexes five typed catalog objects—tables, columns, metrics, relationships, and query history—using object-specific templates rather than raw warehouse rows (Agrawal et al., 23 Jun 2026). TreeCat models catalog metadata as a hierarchy of objects with version ids, distinguishing non-leaf objects, which have a single parent and are uniquely identified by path, from immutable leaf objects, which may have multiple parents and therefore multiple paths. A non-leaf object path has the form

$/oid_1 / oid_2 / \cdots / oid_n$

and is the fundamental identifier for path-based queries and storage ordering (Oh et al., 4 Mar 2025). MatBase pushes the semantic model further inward: its metadata catalog is made of 373 tables and views and stores sets, mappings, constraints, Datalog programs, E-R diagrams, and host-DBMS objects under the (Elementary) Mathematical Data Model and the Entity-Relationship Data Model (Mancas, 9 Apr 2025).

Semantic-artefact repositories use yet another style. OntoPortal-Astro builds on OntoPortal and a rich metadata model derived from MOD, with artefacts imported or rewritten in OWL or SKOS, versioning and mapping support, URI resolution, and FAIRness assessment support through O’FAIRe; artefacts in the prototype are classified as Test, Develop, or Ready (Cecconi et al., 17 Apr 2025).

3. Construction, extraction, and maintenance workflows

A semantic catalog is rarely static. One recurrent design pattern is automated catalog construction from external, heterogeneous, or weakly structured sources. In product synthesis, the end-to-end pipeline has two phases. Offline learning performs web-page attribute extraction and attribute correspondence creation; runtime offer processing then performs schema reconciliation, clustering, and value fusion (Nguyen et al., 2011). The schema reconciliation method uses historical offer-to-product associations, Jensen-Shannon divergence and Jaccard coefficient at merchant/category granularities, and a logistic regression classifier trained on automatically created training sets, so that no manually-labeled data is needed (Nguyen et al., 2011). On Bing Shopping data with 856,781 offers, 1,143 merchants, and 498 catalog categories, the system produced 287,135 synthesized products and 1,126,926 synthesized attribute-value pairs, with attribute precision $= 0.92$ and product precision $= 0.85$ (Nguyen et al., 2011).

Document-derived cataloging follows a parallel logic. CatalogBank converts digitally born engineering PDFs into page images, uses DocumentLabeler as an open-source, offline, privacy-preserving, semi-automatic multimodal annotation tool, and supports PICK, DocBank, XFUND, and FUNSD interchange formats (Bank et al., 2024). The reported release contains 11,984 pages from Misumi, Newark, Thorlabs, McMaster-Carr, 8020, and Grainger, and a PICK baseline on CatalogBank achieved overall $mEP = 0.990$, $mER = 0.994$, $mEF = 0.992$, and $mEA = 0.994$ for layout analysis (Bank et al., 2024). The CED task addresses a more structural problem: recovering the catalog tree from raw document segments. Its TRACER parser operates with four actions—Sub-Heading, Sub-Text, Concat, and Reduce—and on ChCatExt reached Heading F1 $= 90.486$, Text F1 $$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$0, and Overall F1 $$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$1, improving over both a pipeline baseline and a tagging baseline (Zhu et al., 2023).

Recent work extends automation to semantic metadata generation. An LLM study on DCAT-compatible metadata evaluated zero-shot prompting, few-shot prompting, and fine-tuned classification on 9 datasets across 8 DCAT/DCTERMS properties: dcterms:title, dcterms:description, dcterms:creator, dcterms:language, dcterms:spatial, dcat:issued, dcat:keyword, and dcat:theme (Busch et al., 4 Jul 2025). The reported findings are that LLMs can generate metadata comparable to human-created content, few-shot prompting yields better results in most cases, and fine-tuning significantly improves classification accuracy, particularly for dcat:theme (Busch et al., 4 Jul 2025). A plausible implication is that semantic catalogs increasingly combine explicit schemas with automated extraction pipelines rather than relying on manual curation alone.

4. Retrieval substrate, semantic matching, and query execution

A major contemporary shift is the treatment of the semantic catalog as the retrieval target itself. Schema-First Retrieval embeds typed catalog metadata rather than warehouse rows and retrieves across tables, columns, metrics, relationships, and query history using parallel vector search, lineage expansion, cross-encoder reranking, workload memory, and deterministic access-control gates (Agrawal et al., 23 Jun 2026). On CRUSH4SQL, it reached 96.4% table recall@20, and cross-encoder reranking added $$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$2 points at column recall@10; against an equally-templated BM25 baseline, semantic retrieval was $$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$3 points at table recall@5. On SEDE, query history raised table recall@5 from 52.1% to 92.3%. On BIRD, schema-first context reduced SQL execution errors from 15.6% to 6.2%, a 2.5x reduction (Agrawal et al., 23 Jun 2026). Its access-control gate is not a soft prompt instruction but an explicit catalog-level filter:

$$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$4

(Agrawal et al., 23 Jun 2026).

Observability systems adopt an analogous design. A catalog-driven framework for natural-language-to-PromQL uses a hybrid metrics catalog with approximately 1,800 metrics organized into 17 domain categories, augments it with asynchronous runtime discovery of hardware-specific GPU signals, validates it against live Prometheus metrics, and routes queries through intent detection, temporal resolution, category-aware metric selection, and PromQL generation with repair (Sisodia, 15 Mar 2026). The catalog path provides sub-second metric discovery, and the full pipeline completes in approximately 1.1 seconds; this is reported as 3.4× faster than the API fallback path at about 3.7 seconds (Sisodia, 15 Mar 2026).

Catalog query engines can also be specialized systems in their own right. TreeCat compiles a path query into a chain of correlated scan operators, each evaluating one predicate over children of the current context objects and pruning irrelevant branches early (Oh et al., 4 Mar 2025). To support concurrent metadata operations, it introduces a multi-versioned optimistic concurrency control protocol that guarantees serializable isolation. In comparative evaluation, alterTable() on an empty Store Sales table had median latency 8.4 ms and throughput 117 ops/s in TreeCat, versus 180.4 ms and 5.6 ops/s for Iceberg, and 1434.1 ms and 0.6 ops/s for Delta Lake (Oh et al., 4 Mar 2025).

E-commerce systems treat the semantic catalog as a candidate-generation substrate. One model represents each product by seven predefined fields—Title, Description, Product category, Metadata, Brand, Numeric, and Product search terms—encodes them with DistilBERT, and matches queries through a Structured Matching Module that combines $$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$5, $$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$6, and a lexical match matrix $$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$7 (Choi et al., 2020). On PSR, it reached NDCG@1 $$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$8, NDCG@5 $$\{p_1, \ldots, p_n \in P \mid p_i \subseteq C,\; p_i \cap p_j = \emptyset\}, \quad \bigcup_{i=1}^n p_i = C$$9, MAP $/oid_1 / oid_2 / \cdots / oid_n$0, and MRR $/oid_1 / oid_2 / \cdots / oid_n$1 (Choi et al., 2020). At far larger scale, XR-Linear in PECOS casts semantic product search as extreme multi-label classification over 100 million products and reached Recall@100 $/oid_1 / oid_2 / \cdots / oid_n$2 at 1.25 ms/query, a 65% improvement over DSSM (60.9% v.s. 36.8%) (Chang et al., 2021).

5. Governance, trust, quality control, and maturity

Semantic catalogs are also governance systems. XFSC validates each submitted Verifiable Presentation through a three-step pipeline: syntax/format validation, cryptographic verification using URDNA2015 normalization, and schema/semantic verification against the union of all active shapes graphs (Arnold et al., 24 Jan 2025). Only the latest valid version of a subject’s Self-Description becomes active and searchable, and the implementation combines role-based access control, OpenID Connect authentication, JWT-based authorization, and lifecycle states such as active and deprecated (Arnold et al., 24 Jan 2025). This makes trustability an intrinsic property of the catalog rather than an external add-on.

The maturity-model literature makes the same point in broader terms. One assessment framework for catalogues of semantic artefacts identifies 12 dimensions—Metadata, Openness, Quality, Availability, Statistics, PID, Governance, Community, Sustainability, Technology, Transparency, and Assessment—and 43 related features (Corcho et al., 2023). The framework is explicitly feature-based rather than a single aggregate score, and the reported survey of 26 catalogues found strengths in web search GUIs and REST APIs but weaker adoption of standard vocabularies, persistent identifiers, transparency, and formal assessment (Corcho et al., 2023). A common misconception is therefore that semantic catalogs are primarily search interfaces; the maturity model treats governance, curation, sustainability, and preservation as equally constitutive.

Implementation guidance reinforces this broader view. The 5W1H+R work recommends four practical guidelines for catalog implementation: maintain cognitive fit, preserve history, track audit metadata, and support complex metadata values such as JSON (Subramaniam et al., 2021). AICat makes related regulatory requirements explicit: the EU database of high-risk AI systems must be accessible and publicly available with some exceptions, user-friendly, easily navigable, and machine-readable, and AICat is designed to provide consistency, machine-readability, searchability, interoperability, transparency, traceability, and accountability for AI system metadata (Golpayegani et al., 2024).

Security and governance issues also persist inside retrieval-oriented catalogs. Schema-First Retrieval emphasizes that access control must be deterministic and catalog-level, and notes that metadata itself is a security concern because catalog descriptions and query logs can expose sensitive business information even when rows remain protected (Agrawal et al., 23 Jun 2026). This suggests that semantic richness and governance burden typically increase together.

6. Scientific and domain-specific instantiations, misconceptions, and recurring limitations

Astronomy provides several distinct semantic-catalog patterns. Felis acts as the authoritative source from which Rubin Science Platform database schema and TAP-facing metadata are derived, with schema files stored in a Git repository, versioned with Git tags, and validated in GitHub workflows (McCormick et al., 2024). OntoPortal-Astro addresses a different problem: the fragmentation of astronomy, heliophysics, and planetary-science vocabularies across IVOA, SPASE, PDS4, and related ecosystems. It catalogs ontologies, thesauri, vocabularies, taxonomies, and metadata schemas, supports mappings and alignments, and embeds FAIRness assessment support through O’FAIRe and FOOPS! (Cecconi et al., 17 Apr 2025). The Siril HEALpixel Catalog Format represents a third pattern: a compact semantic binary catalog with a fixed 128-byte header, cumulative HEALpixel index, and fixed-length records optimized for fast spatial lookup in Gaia-like catalogs (Knagg-Baugh et al., 21 Jan 2025).

Scientific source catalogs can also have explicitly semantic ambitions. The Big Multi-AGN Catalog (The Big MAC) is described as the first literature-complete catalog of all known confirmed and candidate multi-AGN systems, with 5742 systems drawn from about 600 literature articles, and its schema stores class assignments, confidence flags, selection methods, confirmation methods, coordinates, redshifts, separations, and provenance metadata (Pfeifle et al., 2024). The paper is explicit, however, that the catalog is literature-complete, not astrophysically complete (Pfeifle et al., 2024). The CatSouth and CatGlobe quasar-candidate catalogs show the same tension in a survey context: CatGlobe merges CatSouth with CatNorth into 1,889,813 unique sources at $/oid_1 / oid_2 / \cdots / oid_n$3, but its purity and coverage remain contingent on Gaia DR3 input selection, external multiband surveys, and the classifier and redshift-estimation pipeline (Fu et al., 18 Mar 2025).

Several limitations recur across domains. Schema-First Retrieval states that benchmarks are public proxies rather than real production warehouses, that query history helps only if organizations retain and authorize logs, and that catalog completeness matters because descriptions, sample values, metric definitions, and lineage must be present enough to embed (Agrawal et al., 23 Jun 2026). The PromQL work notes that a catalog is less expressive than a knowledge graph for service dependency or causal relationships, even though it is cheaper to build and maintain and better aligned with zero-configuration startup and fast lookup (Sisodia, 15 Mar 2026). Big MAC, similarly, documents how selection methods such as double-peaked emission lines, periodicity, or X-shaped radio sources can have high false-positive rates (Pfeifle et al., 2024).

A final misconception is that “semantic catalog” always denotes a single layer of abstraction. In practice, the cited work spans catalog mental models, metadata registries, graph-backed trust services, typed retrieval indexes, binary spatial formats, and domain-science object archives. This suggests that the unifying criterion is not implementation style but the catalog’s ability to preserve and operationalize meaning: to make resources discoverable, comparable, validated, queryable, and reusable under explicit semantic structure.