Change Log

August 17, 2026 · View on GitHub

Release v2.2.0 (2026-08-17)

Published NVIDIA NGC artifacts:

Research and reports

  • Routed deep research now uses explicit source-router, structured planner, concurrent researcher, and writer roles, with bounded source-tool batching and no research-plan approval step
  • Report follow-up supports answers over a completed report, child-job cosmetic rewrites, and delta research that carries the parent report forward as context
  • Clarification is more targeted: it can search for context before asking the user to narrow scope or choose an output shape
  • The writer produces figures through an on-demand chart-generation skill in the new visualization skill collection, rendering a sandbox PNG artifact when a sandbox is available and an inline chart spec otherwise; the skill ships enabled only in the skills and sandbox example configs (config_domain_routing_and_skills, config_openshell), while every other config presents chart-worthy data as a Markdown table, and a writer that wants inline charts must be assigned the visualization collection (charts are not sandbox-gated)

Sources and integrations

  • OpenSearch is a first-class knowledge backend for self-hosted, Amazon OpenSearch Service, and Amazon OpenSearch Serverless deployments
  • Azure AI Search is a managed knowledge backend with API-key or Azure identity authentication, namespaced index ownership, and hybrid retrieval
  • Paper search adds SerpAPI and SearchAPI providers alongside Serper; the routed-research profile adds DuckDuckGo news and Polymarket sources
  • You.com adds configurable web search, page-content extraction, cited open-domain research, and finance-focused research tools
  • Nimble adds configurable web search with lite and deep modes, plus an Enterprise-only fast mode and optional focus, country, and locale controls
  • Per-user MCP OAuth adds status, connect, callback, and reconnect flows backed by a token store shared by the API and workers; disconnect and in-worker token refresh are not included
  • A standalone public MCP server exposes stateless submit, poll, and final-report tools over Streamable HTTP with PostgreSQL-backed job state

Sandboxes, artifacts, and policy

  • DeepAgents execution uses a provider-neutral, job-scoped sandbox contract: Modal and OpenShell create one physical sandbox per deep-research job; OpenShell additionally applies and attests the configured policy before execution
  • Opt-in durable artifact capture checkpoints manifest-declared files after successful sandbox execute calls, performs one final manifest-plus-directory scan on success/failure, and preserves earlier checkpoints without delaying cancellation when the provider is busy
  • Captured files store metadata in SQL and bytes in SQL or S3-compatible storage, emit metadata-only artifact.update events for live and replayed Files-tab access, and remain available through job-scoped list/content endpoints that enforce ownership when REQUIRE_AUTH=true
  • Opt-in NeMo Guardrails middleware covers selected workflow and agent input/output boundaries; defining middleware does not activate every boundary
  • Opt-in content encryption protects final async output and selected artifact event content only; it is off by default, forward-only, and does not encrypt checkpoints or most job/event metadata
  • Summary Store database logging masks URL passwords and removes query parameters so credentials and query-string secrets are not written to initialization or lifecycle logs

Deployment and observability

  • The repository source Helm chart honors helm install -n <namespace> for every namespaced resource, including GitOps-rendered deployments; chart metadata advances to aiq2-web 2.2.0 with the aiq 0.0.5 dependency
  • NAT-exported async-job traces preserve configured workflow, task/batch, named-agent, and model/tool hierarchy across concurrent researchers without copying graph-state content into structural agent spans
  • Deep-research intake uses atomic per-principal, deployment-wide, and per-minute admission controls before Dask enqueue; each job also enforces hard input, runtime, plan, report, shared-state, query, note, todo, and source-tool budgets
  • Document ingestion enforces server-side file-count, per-file, aggregate-size, declared-type, and content validation using the same upload settings as the UI
  • The embedded Dask scheduler, dashboard, and worker listeners bind to loopback; production guidance defines the customer-operated security boundaries for external Dask, provider quotas, authentication, and S3-compatible artifact storage

Agent Skills, UX, and developer workflow

  • Consumer Agent Skills now include aiq-deploy and aiq-research; maintainer skills cover workflow configuration, data sources, tools, release QA, PR preparation, prompt/model customization, and CI maintenance
  • The UI surfaces batched researcher activity and improves research-session recovery, expiry handling, and WebSocket delivery reliability
  • Contributor governance and product-level Agent Skill evaluation checks expand release and contribution tooling
  • Pinned to NeMo Agent Toolkit (NAT) v1.8.0

The eleven checked-in workflow configurations are focused profiles; no single profile enables every 2.2 capability.

Release v2.1.0

  • AI-Q REST API with pluggable auth middleware, entry-point-registered token validators, and async job ownership enforcement
  • Auth extensibility hooks (register_token_fetcher, provider lifecycle) and auth refactor eliminating the refresh race
  • Data source registry driving UI toggles, per-message filtering, and agent tool inheritance
  • New exa_web_search data source with full_text and highlights controls
  • Deep researcher consumes DeepAgents skills with a job-scoped Modal sandbox; built-in data-table-analysis skill and configs/config_skills.yml example
  • AI-Q is consumable as a portable Agent Skill (.agents/skills/aiq-research/), with .claude/skills/aiq-research/ retained as a Claude Code compatibility symlink for routed /chat and async job lifecycle against a local AI-Q server
  • Cost analysis tool with pricing configs and profiling example
  • Documented MCP client patterns scoped for 2.1: mcp_client, mcp_service_account, and user-identity tools
  • Prompt restructure across all agents for KV cache prefix reuse
  • Operability: idempotent DB init, tuned Dask/Postgres defaults, request tracing into NAT spans, UI stream-failure hardening
  • New authentication and MCP tools guides; new skills-and-sandbox example
  • Pinned to NeMo Agent Toolkit (NAT) v1.6.0; CVE bumps for Pillow, cryptography, pygments, authlib, pyopenssl, and pytest

Release v2.0.0

Ground-up rewrite of the NVIDIA AI-Q Blueprint, built on the NVIDIA NeMo Agent Toolkit (NAT).

  • Two-tier research architecture with automatic routing between shallow (fast, bounded) and deep (multi-phase, report-grade) research via a single-call Intent Classifier
  • Deep Researcher rebuilt with a three-role subagent architecture (Orchestrator, Planner, Researcher) using the deepagents library, with configurable research loops and per-role LLM assignment
  • New Shallow Researcher agent with tool-call budgets, context compaction, and synthesis anchors for citation-backed answers
  • Clarifier agent with human-in-the-loop plan generation, approval, and feedback before deep research
  • Shallow-to-deep escalation when the shallow researcher detects insufficient results
  • Async Jobs REST API (/v1/jobs/async/) with SSE streaming, event replay, reconnection support, and cooperative cancellation
  • Dask-based distributed execution with configurable workers, heartbeats, and stale job reaping
  • PostgreSQL persistence for job store, event store, LangGraph checkpoints, and document summaries
  • Pluggable Knowledge Layer with factory/registry pattern — swap between LlamaIndex (local ChromaDB) and Foundational RAG (hosted NVIDIA RAG Blueprint) without code changes
  • Multimodal document extraction (VLM-powered image captioning and chart data extraction)
  • Document summaries injected into agent prompts for file-aware research
  • Deterministic citation verification pipeline with five-level URL matching, report sanitization, and audit trail
  • New Next.js frontend with conversational UI, document upload, collection management, and real-time progress streaming
  • Optional OAuth/OIDC authentication with configurable providers
  • Multi-backend observability: Phoenix, LangSmith, W&B Weave, and OpenTelemetry Collector with privacy redaction
  • FreshQA benchmark for shallow researcher factuality evaluation via nat eval
  • Docker Compose and Helm chart deployments with distroless runtime images, non-root execution, and horizontal scaling
  • Native NAT integration — all configuration through YAML with nat run / nat serve / nat eval
  • Four pre-built configs: CLI default, Web + LlamaIndex, Web + Foundational RAG, Hybrid Frontier Model
  • uv workspace monorepo, Jupyter notebook tutorial series, and debug console at /debug
  • Pinned to NeMo Agent Toolkit (NAT) v1.4.0; Python 3.11–3.13; Node.js 22+
  • AI-Q holds top positions on both DeepResearch Bench and DeepResearch Bench II leaderboards (see drb1 and drb2 branches)

Release v1.2.1

  • Upgraded llama-3.3-70b-instruct NIM from version 1.13.1 to 1.14.0
  • Aligned Helm values and referenced Docker image tags with the new nim-llm version
  • Adopted RAG 2.3.2
  • Removed manual NIM_MODEL_PROFILE configuration from Helm values and Docker Compose to rely on automatic profile detection, updated documentation accordingly

Release v1.2.0

  • Added support for Helm deployments
  • Add support and documentation for evaluation
  • Simplified the configuration and integration with RAG, removing nginx
  • Adopted RAG 2.3.0
  • Tested for compatability with RTX Pro 6000

Release v1.1.0

  • Tested for compatability with RAG 2.2.0 release and B200
  • Adds support for NVIDIA Workbench

Release v1.0.0

Initial release of the NVIDIA AI-Q Research Assistant Blueprint featuring:

  • Multi-modal PDF document upload and processing, compatible with the NVIDIA RAG 2.1 blueprint release
  • Demo web application
  • Deep research report writing including human-in-the-loop feedback