Skip to content

ADR-0005: Thin Retrieval Boundary

Target architecture — not the current operating contract

Retrieval is not included in the current release. See its status page.

Status

Accepted — 2026-08-19

Context

Agents and AI Applications need grounded access to enterprise knowledge, citations, freshness, and source permissions. Retrieval implementations can easily expand into source integration, ETL, master data, a warehouse, or an enterprise catalog, creating duplicate authority and an unbounded data-platform roadmap. They can also leak access by treating index presence as permission or copy excessive source content into telemetry and audit stores.

Mission Control needs governed retrieval while keeping enterprise source systems authoritative and the product centered on control and evidence.

Decision

Retrieval is a thin, customer-resident derived index over approved enterprise sources. Every source operation uses a registered Data Connector through the Connector Gateway. Retrieval provides source registration, purpose-specific Collections, governed ingestion and indexing, access-aware keyword, vector or hybrid query, citations, freshness management, source-linked deletion and retention, and retrieval evaluation.

Caller, source, environment, classification, purpose, Collection, and document-level access are enforced at ingestion and query time. Index presence never grants access. Mission Control stores only approved derived content, embeddings, citation and ACL metadata, deletion markers, and freshness state. Run evidence retains source and citation references rather than unnecessary source payload copies.

Retrieval does not become ETL, reverse ETL, master-data management, a warehouse or lakehouse, an enterprise data catalog, or a source-of-record replacement.

Consequences

  • Agents and Applications receive governed passages with citations while customer systems remain authoritative.
  • Retrieval reuses MCP Data Connectors, workload identity, policy, lifecycle, evidence, evaluation, and incident contracts instead of creating a private source integration layer.
  • Permission revocation, deletion, retention expiry, or source retirement must propagate into query behavior and derived content.
  • Retrieval quality includes relevance, coverage, citation correctness, freshness, leakage prevention, and access-control behavior.
  • The product accepts narrower ingestion and transformation scope than a general data platform; customers use existing data systems for broad movement, normalization, governance, and mastering.
  • Embedding and reranking inference, when used, still resolves an approved Virtual Model through the Model Gateway.
  • Future features must remain traceable to governed retrieval and cannot use convenience as a reason to absorb general enterprise data-platform responsibilities.

Current governance mapping

The current source release uses Organizations, Teams, real user/service-account virtual keys and separate human management roles. Agent is managed application content; deployments own secretless execution access and publisher funding. Observe uses Requests, Traces, Sessions and Usage with separate metadata/content/export permissions. Historical identity terms in this target proposal are not the operating contract. See Current release.