Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 33 additions & 2 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -1,12 +1,43 @@
# Changelog

## [Unreleased]
## [v3.6.0] - 2026-07-19

### User and operator impact

Adds opt-in provisioning for generic, approved, preconfigured MCP Server knowledge sources, resolving [#567](https://github.com/Azure/GPT-RAG/issues/567) through [GPT-RAG PR #568](https://github.com/Azure/GPT-RAG/pull/568), [docs PR #569](https://github.com/Azure/GPT-RAG/pull/569), and [orchestrator PR #275](https://github.com/Azure/gpt-rag-orchestrator/pull/275). The feature is **Preview and disabled by default**, so existing deployments do not change unless an operator explicitly enables it.

Provisioning from GPT-RAG `v3.6.0` and [orchestrator `v3.7.0`](https://github.com/Azure/gpt-rag-orchestrator/releases/tag/v3.7.0) must be deployed together before enabling MCP. The configuration governs trusted HTTPS endpoints, exact tool allowlists, output parsing, per-source failure behavior, reasoning effort, runtime/output limits, and request-time authentication metadata.

### Added

- **Generic MCP Server knowledge source template.** Foundry IQ can call allowed tools on one or more operator-approved MCP servers and combine those results with existing knowledge sources. The Search integration uses API `2026-05-01-preview`; Azure Monitor MCP over workspace-based Application Insights is a worked example, not a special product code path.
- **Fail-closed source validation and canonical configuration.** Provisioning rejects malformed or empty source definitions, duplicate names, untrusted or unsafe endpoints, unsupported tool/output controls, forbidden headers, literal credentials, and invalid runtime limits before any App Configuration or Search write. Disabled configuration remains canonical and byte-identical to the prior rendered behavior.
- **Secure request-time authentication metadata.** Managed identity, OBO, Key Vault secret-name references, and no-auth mode are supported without rendering credential values into Search resources. Operators must apply least privilege and keep secrets out of source JSON.
- **Operator controls and rollback.** `FOUNDRY_IQ_MCP_ENABLED=false` remains the default and rollback gate. Trusted-host and tool allowlists, bounded reasoning/runtime/output, explicit parsing, and per-source failure behavior constrain the preview integration.

### Changed

### Fixed
- **`manifest.json` now identifies umbrella release `v3.6.0` and pins orchestrator `v3.7.0`.** UI `v2.3.13`, ingestion `v2.4.14`, and infra `v2.3.0` are unchanged.

### Security and operational guidance

MCP tool safety does not guarantee that model-generated arguments are semantically correct. Use only trusted endpoints, exact read-only tool allowlists, bounded queries, least-privilege identities, and auditable server-side enforcement. The Azure Monitor example must retain workspace scope, bounded time ranges and row counts, and generated-KQL review. Production enablement requires a canary against the operator's approved endpoint and credentials.

To roll back, 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 be deleted separately after confirming that no knowledge base references it.

### Validation

| 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 |

- 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 MCP remains Preview, disabled by default, fail-closed, fully regression-tested, and reversible; production enablement still requires an operator-owned canary.

## [v3.5.1] - 2026-07-15

Expand Down
Loading
Loading