mdh-reference

Master Data Hub (MDH) Reference

This skill provides comprehensive reference documentation for Rossum's Master Data Hub — the matching engine that connects extracted document data to master data records. It covers the MDH management API, the hook configuration model, and query design patterns.

Reference files:

• mdh-api-reference.md — MDH API endpoints (dataset CRUD, fuzzy search, operation status, Search Indexes V2) and hook configuration schema (MatchConfig, DMDatasetSource, Mapping, ResultActions, query cascade model)
• mdh-matching-queries.md — Query design rules (DO/DON'T), score normalization patterns, unique-result filters, GL coding dropdown pre-selection, real-world examples (supplier matching, PO line items, delivery address resolution), and Atlas Search index recommendations

Use this knowledge when:

• Building MDH matching configurations (hook JSON with source, mapping, result_actions)
• Designing query cascades (exact → fuzzy → fallback)
• Using score normalization or $setWindowFields unique-result patterns
• Configuring memorization via $unionWith patterns
• Managing MDH datasets via the API (upload, replace, update, delete)
• Enabling fuzzy search on datasets
• Declaring or debugging Atlas Search indexes (Search Indexes V2 — a dataset subresource, not a Data Storage call)
• Debugging matching hook configurations

Mandatory pre-flight for any $search query. Before shipping or debugging a $search stage, list the search indexes on the dataset (data_storage_list_search_indexes) and confirm the named index is declared, is queryable: true, and maps every field you reference. A misnamed index, or one missing a mapping, returns zero results silently with status: completed — the hardest MDH failure mode to diagnose. Full recipe in mdh-matching-queries.md → Atlas Search Index Pre-flight.

Declared indexes are durable (MDH reconciles the registry onto the collection), so do not write index-repair loops and do not treat a missing index as drift — it is a missing declaration. Manage them through the Search Indexes V2 API; btree indexes are a separate, non-self-healing concern.

When writing or debugging queries, if the dataset structure is not already known from context, try to discover it using the rossum-api MCP tools: data_storage_list_collections to find available collections and data_storage_aggregate with [{"$sample": {"size": 1}}] to retrieve a sample record. Fall back to asking the user only if the MCP tools are not available.