v2.2.1

NVIDIA AI-Q v2.2.1

AI-Q 2.2.1 is a focused maintenance release that improves runtime correctness, release packaging, citation handling, model output capacity, and documentation accuracy.

Highlights

• More reliable asynchronous jobs: Adds readiness checks before submission, prevents invalid jobs after intent-provider failures, and avoids retrying authentication failures.
• Stronger release packaging: Restores required runtime prompt assets and validates all 11 shipped workflows from an installed, VCS-less release artifact.
• Configurable citation enforcement: Shallow research returns the sanitized generated answer by default while retaining strict fail-closed behavior with enforce_citations: true.
• Expanded Lightning output capacity: Raises the Nemotron 3.5 Lightning shallow-research output limit to 32,768 tokens across shipped default configurations.
• Safer connector output: Improves web-search content escaping and citation handling for escaped URLs and titles.
• Security and documentation maintenance: Updates the MCP production lock for the latest pip security advisory and corrects stale references across deployment, integration, profiling, architecture, and contributor documentation.

Published artifacts

Artifact	Versioned reference
Backend container	nvcr.io/nvidia/blueprint/aiq-agent:2.2.1
Frontend container	nvcr.io/nvidia/blueprint/aiq-frontend:2.2.1
Helm chart	nvidia/blueprint/aiq2-web:2.2.1

Configuration notes

• Shallow citation enforcement remains backward-compatible by default. Set enforce_citations: true to require strict fail-closed citation validation.
• The 32,768-token Lightning output limit applies to shipped default profiles. Benchmark profiles intentionally remain at 8,192 tokens.
• This patch release contains release-scoped fixes and does not introduce develop-only AI-Q 3.0 features.
• The Brev getting-started notebook follows the maintained release/2.2 branch.

What's Changed

Runtime reliability and packaging

• #460 Restore release packaging and harden connector, asynchronous-job, and intent-provider behavior by @KyleZheng1284
  Backports #454 and #455.

Research and documentation

• #462 Add configurable citation enforcement, raise the Lightning shallow output limit, update the MCP production lock, and correct stale documentation references by @AjayThorve
  Backports #456 and #461, with original contributions from @shubhadeepd and documentation corrections from @thanhpt1110.

Deployment and release

• #451 Update the Brev notebook repository and architecture references for the maintained release branch by @soumilinandi
• #463 Prepare the AI-Q 2.2.1 Python packages, Compose deployment, Helm chart, documentation metadata, SBOM expectations, and release validation by @AjayThorve

Release stats

• 4 merged release PRs
• 5 release-branch commits
• 87 files changed
• 2,138 additions and 295 deletions

Learn more

• Detailed release changelog
• Full changelog: v2.2.0...v2.2.1
• AI-Q documentation

v2.2.1-rc1

What's Changed

• fix(brev): brev notebook repo clone update to release tag by @soumilinandi in #451
• fix: backport/2.2.1 remediation by @KyleZheng1284 in #460
• fix: backport v2.2.1 citation, output, and docs follow-ups by @AjayThorve in #462

Full Changelog: v2.2.0...v2.2.1-rc1

v2.2.0

NVIDIA AI-Q v2.2.0

AI-Q 2.2 expands the blueprint into a more extensible, production-ready platform for enterprise research workflows.

Highlights

• New role-specific OSS model defaults: The default AI-Q profiles now pair NVIDIA Nemotron 3.5 Lightning for intent classification and shallow research with Nemotron 3 Ultra for clarification and every deep-research role—combining fast routing and bounded answers with stronger multi-step research and report generation.
• More capable research workflows: Routed deep research now uses explicit source-routing, planning, concurrent research, and report-writing roles. Report follow-ups support questions, rewrites, and additional research grounded in a completed report.
• Broader enterprise connectivity: First-class OpenSearch and Azure AI Search backends, expanded paper-search providers, You.com and Nimble integrations, per-user MCP OAuth, and a public AI-Q MCP server.
• Sandboxed, durable execution: Job-scoped Modal and OpenShell sandboxes, OpenShell policy attestation, checkpointed artifact capture, and SQL or S3-compatible artifact storage.
• Stronger security boundaries: Optional workflow and agent guardrails, scoped async-content encryption, protected document ingestion, secret-safe logging, and bounded research workloads.
• Improved operability: Admission controls, preserved workflow and agent trace hierarchy, hardened Dask defaults, and more reliable async-job lifecycle handling.
• Better UX and developer workflow: Improved session recovery, batched researcher activity, WebSocket reliability, consumer and maintainer Agent Skills, and expanded evaluation tooling.
• NeMo Agent Toolkit: Updated to NAT v1.8.0.

Published artifacts

Artifact	Versioned reference
Backend container	nvcr.io/nvidia/blueprint/aiq-agent:2.2.0
Frontend container	nvcr.io/nvidia/blueprint/aiq-frontend:2.2.0
Helm chart	nvidia/blueprint/aiq2-web:2.2.0

Configuration notes

• The checked-in configurations are focused profiles; no single profile enables every AI-Q 2.2 capability.
• Per-user MCP OAuth includes status, connect, callback, and reconnect flows. Disconnect and in-worker token refresh are not included.
• Async-content encryption is opt-in and scoped to final output and selected artifact-event content; checkpoints and most job metadata are outside its scope.

What's Changed

Research and reports

• #245 Clarifier searches before asking clarification questions by @torkian
• #267 Make the deep researcher customizable, flexible, and less token-expensive by @cdgamarose-nv
• #271 Add report follow-up jobs and chat routing by @AjayThorve
• #286 Return deep-research results through ResearchNotes by @zhongxuanwang-nv
• #289 Salvage inline reports and persist planner plans by @KyleZheng1284
• #303 Remove plan approval while retaining context and output clarification by @ashan-nv
• #306 Recognize sandbox sources during deep-research routing by @KyleZheng1284
• #311 Clarify todo-writing instructions by @cdgamarose-nv
• #313 Add report follow-up skills by @peterychang
• #318 Reap running jobs that never stored an event by @torkian
• #347 Safely configure asynchronous deep research by @KyleZheng1284
• #360 Handle interrupted jobs and the JSON escalation contract by @tanleach
• #375 Require writer-committed final reports by @KyleZheng1284
• #385 Honor async-submit tool and agent availability by @tanleach
• #400 Update research configurations by @soumilinandi
• #402 Harden the default guardrails configuration by @soumilinandi
• #405 Improve empty-source outcomes and artifact documentation by @tanleach
• #413 Support GPT frontier structured responses by @AjayThorve
• #414 Update default model successors by @AjayThorve
• #422 Stabilize the research runtime by @AjayThorve
• #429 Resolve AI-Q 2.2 release-readiness findings by @AjayThorve
• #430 Enable Nemotron 3.5 Lightning defaults by @AjayThorve
• #433 Preserve GPT frontier model roles by @AjayThorve
• #439 Enforce shallow-research citation integrity by @AjayThorve
• #446 Document hosted Lightning limitations and harden Brev shallow research by @AjayThorve

Sources and integrations

• #224 Support data-source selection during job submission by @tanleach
• #226 Add headless-mode support to the chat endpoint by @cdgamarose-nv
• #228 Add OpenSearch as a first-class knowledge backend by @rhossi
• #261 Add Nimble web search by @wildcard
• #266 Add configurable Tavily base URL support by @cdgamarose-nv
• #276 Add per-user MCP authentication by @ashan-nv
• #278 Add SerpAPI and SearchAPI paper-search providers by @gcasilva
• #308 Add the Azure AI Search knowledge backend by @harmke
• #316 Add the You.com API suite by @eddy-nassif
• #319 Add the public AI-Q MCP server by @tanleach
• #336 Stabilize the Azure AI Search file lifecycle by @harmke
• #361 Upgrade the MCP SDK for CVE-2026-59950 by @tanleach
• #367 Propagate REST submission tokens to asynchronous workers by @AjayThorve
• #407 Use source metadata descriptions in the knowledge layer by @AjayThorve

Sandboxes, artifacts, and security

• #280 Add provider-neutral sandboxes and a durable artifact runtime by @KyleZheng1284
• #291 Encrypt asynchronous job content artifacts by @tanleach
• #292 Add workflow and agent boundary guardrails by @ericevans-nv
• #298 Isolate and attest OpenShell jobs by @KyleZheng1284
• #300 Add S3-compatible artifact storage by @soumilinandi
• #307 Redact secrets from summary database logs by @harmke
• #314 Persist and surface sandbox artifacts by @KyleZheng1284
• #358 Restore PII masking and contain guardrail output failures by @tanleach
• #372 Enable Modal artifact capture by @KyleZheng1284
• #377 Upgrade the certified OpenShell stack by @KyleZheng1284
• #393 Upgrade vulnerable dependencies for AI-Q 2.2 by @tanleach
• #401 Harden release security boundaries by @AjayThorve

Deployment and observability

• #309 Make the Helm chart honor the selected release namespace by @e-minguez
• #320 Surface batched researcher activity by @AjayThorve
• #321 Preserve AI-Q agent trace hierarchy by @AjayThorve
• #426 Pin the deployment Python runtime to 3.13.14 by @tanleach

Agent Skills and UX

• #164 Correct session-title updates by @DinaLaptii
• #165 Open the data-source panel by default by @DinaLaptii
• #171 Replace the settings panel with a popover by ...

v2.2.0-rc9

What's Changed

• fix: resolve VDR 2.2 release-readiness issues by @AjayThorve in #429
• feat: enable Nemotron 3.5 Lightning defaults by @AjayThorve in #430

Full Changelog: v2.2.0-rc8...v2.2.0-rc9

v2.2.0-rc8

What's Changed

• fix(deploy): pin Python runtime to 3.13.14 by @tanleach in #426

Full Changelog: v2.2.0-rc7...v2.2.0-rc8

v2.2.0-rc7

What's Changed

• feat: harden release 2.2 security boundaries by @AjayThorve in #401
• fix(ci): unblock release checks by @AjayThorve in #415
• support GPT frontier structured responses by @AjayThorve in #413
• feat(brev): Update Brev getting started notebook by @soumilinandi in #412
• fix: update available default model successors by @AjayThorve in #414
• fix: stabilize research runtime by @AjayThorve in #422

Full Changelog: v2.2.0-rc6...v2.2.0-rc7

v2.2.0-rc6

What's Changed

• docs(sandbox): clarify OpenShell job isolation by @KyleZheng1284 in #387
• fix(security): upgrade vulnerable dependencies for release 2.2 by @tanleach in #393
• docs: clarify knowledge collection routing by @KyleZheng1284 in #391
• docs: refresh AI-Q 2.2 architecture by @AjayThorve in #395
• fix(research): update configs by @soumilinandi in #400
• fix(research): harden default guardrails config by @soumilinandi in #402
• fix(research): improve empty-source outcomes and artifact docs by @tanleach in #405
• fix(knowledge_layer): use source metadata description by @AjayThorve in #407

Full Changelog: v2.2.0-rc5...v2.2.0-rc6

v2.2.0-rc5

What's Changed

• fix(openshell): upgrade certified stack to 0.0.88 by @KyleZheng1284 in #377
• chore(release): prepare AI-Q 2.2.0 by @tanleach in #379
• fix(api): honor async submit tool and agent availability by @tanleach in #385

Full Changelog: v2.2.0-rc4...v2.2.0-rc5

v2.2.0-rc4

What's Changed

• feat(docs): improve skill and documentation by @soumilinandi in #374
• fix(config): enable Modal artifact capture by @KyleZheng1284 in #372
• fix(deep-research): require writer-committed final reports by @KyleZheng1284 in #375

Full Changelog: v2.2.0-rc3...v2.2.0-rc4

v2.2.0-rc3

What's Changed

• fix(mcp): bump MCP SDK to 1.28.1 for CVE-2026-59950 by @tanleach in #361
• fix(guardrails): restore PII masking and contain output failures by @tanleach in #358
• docs: fix v2.2.0 documentation defects by @tanleach in #359
• fix(aiq-research): handle interrupted jobs and JSON escalation contract by @tanleach in #360
• fix(jobs): reap RUNNING jobs that never stored an event by @torkian in #318
• fix(auth): propagate REST submit token to async workers by @AjayThorve in #367
• fix(knowledge): stabilize Azure Search file lifecycle by @harmke in #336

Full Changelog: v2.2.0-rc2...v2.2.0-rc3