What's new, by capability.

Preview

DAMA-DMBOK2: requirements with evidence

A new governance cockpit connects knowledge areas, requirements, work items and maturity assessments. Owners can record evidence for progress and see which requirements still need validation.

Preview

Accountability for delegated actions

Delegated actions carry the principal’s roles and domain scope. Purpose and checkpoint records make it possible to review who authorised an action and on whose behalf it ran.

Preview

Processing register and replayable local demo

A processing register and local collector demo link processing events to their context. The demo uses fixture replay to inspect the registration path; it does not establish complete operational coverage.

Preview

One metric definition across data sources

Mensura compiles governed metric definitions into SQL for different data platforms. Model validation, execution interfaces and policy checks connect the analytical definition to the organisation’s ontology.

Preview

KPI answers with metric citations

Ask Next can call Mensura and attach metric citations to its answer. Provenance and trace identifiers connect a reported number to the metric definition and execution that produced it.

Evaluation tooling

Governed text-to-SQL evaluation

Evaluation tooling compares a governed LLM-to-OBSQL path with directly generated SQL. It measures execution outcomes, query-object agreement, compile errors and correct refusal on unanswerable questions.

Preview

Search and grounded answers through MCP

Search and answer tools can be reused by MCP clients and delegated agents. Tool scopes, workspace bindings, authentication and rate limits make the permitted knowledge and operations explicit.

Preview

Answer dossiers for reviewers

Answer dossiers bring sources, grounding outcomes, metric citations and execution traces together. Reviewers can examine the evidence behind a response and distinguish a blocked answer from a technical failure.

Evaluation tooling

Stable goldsets and regression evaluation

Evaluation now supports stable evidence identifiers, holdout splits, citation precision, correct abstention and multi-turn conversations. Feedback supplies candidates for human curation. Candidate checks currently report warnings rather than enforce a blocking release gate.

Retrieval measurements by question type

The August devtest runs compare BM25, hybrid retrieval and reranking on a synthetic set. Overall nDCG@10 was 0.498, 0.562 and 0.711 respectively. These admin-scope measurements do not establish user-level answer quality; query-embedding time is excluded.

Faster SPARQL planning and index access

Cardinality-aware joins, more selective index access and bounded sorting reduce work for specific SPARQL query shapes. Before-and-after measurements include result hashes, so performance can be examined alongside equivalent results.

A reproducible Apache Jena comparison

In a July single-node CRC test capped at 2 CPU and 4 GiB per engine, IndentiaGraph achieved 1.62× and 3.10× Jena’s throughput on the RDF 1.1 and RDF 1.2 query mixes. The small synthetic workload does not measure HA scaling; Turtle 1.2 reifier loading via Graph Store failed for IndentiaGraph in this run.

Compaction, replica inspection and repair

Storage work includes compressed SSTables, bounded compaction output, replica audits and administrative repair of corrupt quads. Cluster changes include online shard splitting; supported deployment modes and upgrade paths need environment-specific validation.

Synonyms and warm-index refresh

Index-time synonym expansion extends text retrieval. Warm text indexes can be refreshed without blocking queries, adding to the previously released facets and fuzzy matching.

StorageMap and integrity inspection

Pharos adds a reusable storage-distribution view, integrity inspection, administrative quad-repair controls and write metrics. Operators can relate storage use to the database’s current health.

Preview

NL-SBB, MIM and DCAT-AP-NL support

New work connects concept import and export, profile validation, MIM RDF import and catalogue exchange. Profile validation, internal quality checks and substantive steward review remain separate outcomes.

Entity resolution and domain classification

Lexicon and the extraction chain extend identity resolution, process-domain classification and sense definitions on asset links. Model-based classification requires independently annotated examples before making accuracy claims.

Preview

Domain packs from model to governed application

Domain packs connect shared vocabulary, model artefacts, APIs and governed writes. The metadata model, the executable interface and the materialised data are tracked as distinct stages.

More standards in the Loom marketplace

The marketplace adds Geonovum, DPV and DPROD entries for discovering and selecting semantic standards. Catalogue presence does not by itself establish full implementation or conformity.

Embedded-image OCR with execution evidence

DOCX extraction can request OCR for embedded pictures. The parsing chain records picture OCR only when execution evidence is present, preserves text hashes and carries the original container path into extracted Markdown.

Traceability across bronze, silver and gold

Layered processing records source and transformation context, including silver snapshot references and content hashes for gold runs. A derived result can be reviewed against the inputs used to produce it.

Preview

Markeer connects source work and modelling

Document, ontology and bulk workspaces are brought together. Selected source signals can become an ontology proposal on a controlled change branch, retaining the link to the material being reviewed.

Preview

Device indexing connected to Enterprise Search

Device content is exposed through the indexing path rather than client-side polling. Scans are visible as Backoffice runs and endpoint files receive a user-bound ACL. Binary availability remains specific to the release and platform.

Memory proposals and source-linked review

Cortex and Memex work connects source analysis, stored knowledge and acceptance of proposals. Memory access is tied to service identity and review policy rather than treated as an unrestricted shared store.

Preview

From architecture design to engineering handoff

Graph Studio adds C4 views, a promotion cycle and an engineering handoff from a selected design. ArchiMate profile checks describe internal test coverage, not an external certification.

Delegation with bounded scope and recoverable quests

The Agent Orchestrator carries principal roles and domain scope into delegated runs, routes workloads to shared MCP services and reconciles queued quests whose eligibility event was missed.

Reviewable enrichment and local authoring assets

Enrichment flows connect terminology proposals to Backoffice review. Studio can serve its code editor and map tiles locally in container mode, reducing reliance on external requests during authoring.

Government Q&A inline

Parliamentary questions and official decisions are now rendered directly inside search results: question and answer visible without leaving the interface. Linked to the responsible person's profile, the originating dossier and the source document.

Person profiles with linked documents

Searching for a minister or official now surfaces a profile panel with all associated parliamentary contributions, publications and decisions, grouped by date and type.

Shareable deep-links for searches

Filter state is now stored in the URL. A search with active source and date filters produces a link that opens the exact same result set for anyone with access.

Term explanation panel

Recognised terms in a search query trigger an inline explanation panel drawn from the organisation's own terminology graph, without leaving the search results.

Configurable per-source timeouts

Each data source can now have its own query timeout, preventing a slow connector from blocking results from faster sources.

Response cache without external dependencies

Search results are now cached in-process. Repeated queries return in milliseconds without requiring Redis or any other external caching service.

Intelligent multi-parser routing

The platform now selects the best document extraction engine automatically for each file type, using a self-learning routing model that improves with every conversion. Multiple engines race in parallel; the highest-quality result wins.

LLM quality judge

Extraction results are evaluated by a language model that scores layout fidelity, text completeness and structure accuracy, and triggers a re-extraction via a different engine if the threshold is not met.

Expanded file type support

The document conversion pipeline now accepts all major office formats, scanned PDFs, images, spreadsheets and presentations through a single upload interface.

Bilingual conversion wizard

The document conversion interface is now fully available in both Dutch and English, with a shared language switcher in the header.

Goal-aware extraction modes

Extraction can now be configured per job: RAG mode prioritises clean, chunked text output; extraction mode preserves tables, structure and layout metadata.

New product: Lexicon

Lexicon is a dedicated product for managing the organisation's business vocabulary. Terms are proposed by users, reviewed by stewards and formally approved (or deprecated) by domain owners.

Terminology lifecycle

Each term has a full lifecycle: draft → under review → approved → deprecated. Every status change is logged with actor, timestamp and rationale.

SHACL validation

Approved terms are validated against the organisation's SHACL shapes before publication, ensuring structural consistency with the business ontology.

Gap detection

Lexicon analyses incoming documents and queries against the approved vocabulary and surfaces terms that appear frequently but have no approved definition yet.

DataHub integration

The approved vocabulary is automatically synchronised to DataHub as a governed glossary, making terminology available across the data catalogue without manual duplication.

Usage analytics

Each term tracks where it is used, in which documents, queries and agent responses, giving stewards a clear picture of vocabulary reach and adoption.

SPARQL 1.2 federated queries

IndentiaDB now supports the W3C SPARQL 1.2 specification, including the SERVICE directive for cross-graph federation. A single query can now span multiple named graphs across tenant boundaries under controlled access.

RDF 1.2 compliance

Support for RDF 1.2 triple terms (reification), native directional language strings and a full W3C conformance harness. Indentia is now compliant with the latest W3C graph standards.

Production-grade durability

The write-ahead log now uses fsync on every commit, with CRC32 checksums per record and automatic torn-tail repair on startup. A committed write is guaranteed to survive a power loss, kernel panic or unclean shutdown. No data is ever partially written.

Full-text search with facet aggregations

The integrated full-text engine now supports field-level facet aggregations with year histograms alongside the existing keyword and fuzzy search, enabling drill-down filtering without a separate search engine.

Native fuzzy matching

Fuzzy full-text queries now use a native Damerau-Levenshtein engine that correctly handles transpositions and adjacent character swaps, the most common typo patterns in Dutch.

Hard query memory limits

Each query now runs under a configurable memory ceiling with a background watchdog. Queries that exceed the limit are terminated cleanly rather than causing OOM pressure on the host.

Quest board

The Agent Orchestrator now includes a quest board: a structured task management interface where agents receive assignments, track progress and report completion, with full audit history per quest.

WebSocket terminal proxy

Operators can now connect a live terminal session to any running agent sandbox over WebSocket, enabling real-time inspection and debugging without breaking sandbox isolation.

Multi-tenant identity bootstrapper

Each tenant now gets a dedicated ServiceAccount with scoped credentials, provisioned automatically at tenant creation. OPA policies govern what each agent identity is permitted to do within its tenant boundary.

MCP server configuration per session

Each agent session can now load a specific set of MCP (Model Context Protocol) server connections from configuration, making the set of available tools explicit and auditable per run.

Rijksoverheid.nl connector

A new connector ingests publications, news items and policy documents from Rijksoverheid.nl with bitemporal semantics, meaning you can query the platform's view of the government website at any historical point in time.

Tweede Kamer documents

Parliamentary papers and Kamerstukken are now ingested directly in binary format, parsed by the document intelligence pipeline and indexed with full provenance metadata.

Source specification registry

All data source connectors now declare a machine-readable specification: schema, capabilities, supported extraction modes and quality contract. This registry is used by the platform to route data and by teams to onboard new sources consistently.

Cooperative backpressure

Connectors can now signal the ingestion pipeline to pause or slow down when upstream systems are under load, preventing connectors from overloading source APIs during peak hours.

RobBERT-2022 embedding service

The platform now ships a dedicated Dutch embedding model based on RobBERT-2022 (768-dimensional). This model understands Dutch morphology, compound words and domain vocabulary significantly better than translated or multilingual alternatives.