Skip to content

v3.6.0

Choose a tag to compare

View all tags
@placerda placerda released this 19 Jul 15:43
v3.6.0
a10b42f
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Changed

  • Adds opt-in provisioning for generic, approved, preconfigured MCP Server knowledge sources, resolving issue #567 through GPT-RAG PR #568, docs PR #569, and orchestrator PR #275.
  • The capability is Preview and disabled by default. Existing deployments do not change unless an operator explicitly enables it.
  • GPT-RAG v3.6.0 provisioning and orchestrator v3.7.0 must be deployed together before enabling MCP. Neither component alone implements the complete contract.
  • Operators can govern trusted HTTPS endpoints, exact tool allowlists, output parsing, per-source failure behavior, reasoning effort, maximum runtime, output limits, and request-time authentication metadata.
  • The live Generic MCP server how-to covers configuration, trust boundaries, rollout, canary validation, and rollback.

Component versions

Component Version
gpt-rag-ui v2.3.13
gpt-rag-orchestrator v3.7.0
gpt-rag-ingestion v2.4.14
infra / AI Landing Zone v2.3.0

Security and authentication

Provisioning fails closed before any App Configuration or Search write when source JSON is malformed or empty, names collide, an endpoint is not an explicitly trusted HTTPS host, tool or output controls are invalid, runtime limits are unsafe, forbidden headers are used, or credential-shaped values are embedded.

Managed identity, OBO, Key Vault secret-name references, and no-auth mode are supported as request-time metadata. Credential values are not rendered into Search resources. Use least-privilege identities, exact read-only tool allowlists, bounded operations, and auditable server-side enforcement. MCP tool safety does not guarantee that model-generated arguments are semantically correct.

Azure Monitor worked example

Azure Monitor MCP over workspace-based Application Insights is an example, not a product-specific code path. Keep it workspace-scoped, limit time ranges and row counts, review generated KQL, and confirm the current endpoint tool name through tools/list instead of hard-coding a version-dependent name.

Rollback

Set FOUNDRY_IQ_MCP_ENABLED=false and rerun post-provisioning or provisioning. This removes MCP references from the knowledge base and restores disabled planning behavior. An unused top-level Search knowledge-source resource may remain and should be deleted separately only after confirming that no knowledge base references it.

Validation

  • GPT-RAG configuration and template suites: 109 tests passed, 66 subtests passed.
  • Orchestrator v3.7.0: 457 tests passed, 65 subtests passed.
  • Python compilation, PowerShell AST/preflight, JSON/manifest/remote-tag checks, canonical fixture compatibility, disabled-byte identity, Bicep build/size gate, strict documentation build, and Pages publication passed.
  • No live Azure MCP end-to-end test was run because no approved MCP endpoint or credentials were available. The release is acceptable because the capability remains Preview, disabled by default, fail-closed, regression-tested, and reversible. Production enablement still requires an operator-owned canary with its approved endpoint and credentials.
Assets 2
Loading