mcp-data-platform-v1.91.0

DataHub context documents are now discoverable in search (#692)

Context documents (DataHub's home for runbooks, FAQs, and reference guides linked to data assets) are now a first-class search source, at full parity with knowledge pages. Before this release they were undiscoverable from unified search: the catalog source only searched datasets, so the knowledge in them was stranded.

The new context_documents source has two arms:

• By relevance: search with an intent ranks context documents by relevance.
• By entity: search with entity_urns returns the documents linked to those entities (lineage-expanded), so a runbook attached to a dataset surfaces when you search that dataset's URN, exactly as a knowledge page that references it does.

Results from both arms merge and de-duplicate by document URN, with the entity match ranked first.

Visibility and publication follow the DataHub contract: the relevance (global search) arm returns only globally-visible, published documents, while the entity arm additionally surfaces documents marked "accessible only through linked assets" (the entity lookup is that path). Drafts are excluded from both.

Search source names are now by knowledge type (breaking)

Two sources were renamed so every source reads as the kind of knowledge it holds rather than the backend or an over-broad noun:

• datahub → catalog (the technical data catalog: datasets, columns, schemas, lineage)
• documents → context_documents (curated documents about catalog entities)

Breaking change. If you pass a sources filter to search, use catalog instead of datahub. The result provenance label on catalog hits also changes from datahub to catalog. The DataHub connection name and the datahub_* tool names are unchanged; only the user-facing search-source identity moved off the backend name.

The full source set is now: catalog, context_documents, knowledge_pages, memory, insights, feedback, assets, prompts, endpoints, connections.

Other search improvements

• Unknown source names are reported. A sources entry that matches no known source (a typo) is echoed back in a new unknown_sources field on both the MCP search tool and the portal GET /search response, instead of silently returning nothing.
• Concurrent fan-out. The knowledge router now queries its independent providers concurrently rather than serially, with per-provider failure isolation (one unhealthy source no longer blanks the search) and deterministic, registration-ordered results.
• Full DataHub search observability. All DataHub search-family calls (catalog search, semantic search, document search, and related-document lookup) now record request metrics and a trace span.

Dependency

• Requires github.com/txn2/mcp-datahub v1.10.1, which adds SearchDocuments and the full related-document projection that back the two search arms. Context documents require DataHub 1.4.x or later.

Installation

Homebrew (macOS)

brew install txn2/tap/mcp-data-platform

Claude Code CLI

claude mcp add mcp-data-platform -- mcp-data-platform

Docker

docker pull ghcr.io/txn2/mcp-data-platform:v1.91.0

Verification

All release artifacts are signed with Cosign. Verify with:

cosign verify-blob --bundle mcp-data-platform_1.91.0_linux_amd64.tar.gz.sigstore.json \
  mcp-data-platform_1.91.0_linux_amd64.tar.gz