Agentic Systems

AI-Q Research Assistant (AIQ Blueprint)

Vendor-stack study notes — verified against NVIDIA's own pages on the access date shown; product names and versions move fast, re-verify before relying on them.

Accessed/verified: 2026-07-14. Repo: github.com/NVIDIA-AI-Blueprints/aiq, current release v2.1.0 (2026-05-19) [aiq-github].

What it is. "An open reference example for building intelligent AI agents that connect to your enterprise data, reason using state-of-the-art models, and deliver trusted business insights" [aiq-github] — concretely, a deep-research assistant that serves both quick cited answers and long-form report-style research from one system. Where the RAG blueprint is the fixed pipeline, AI-Q is the agentic layer on top: retrieval becomes a tool inside a planned, multi-agent loop (the fixed-pipeline-vs-loop trade is argued in the enterprise-RAG brief §3).

Multi-agent architecture

Built on the NVIDIA NeMo Agent Toolkit v1.8.0 plus LangChain Deep Agents (>= 0.6.5), as a LangGraph state machine [aiq-github]. The agents:

  • Orchestration node — classifies intent (meta vs research), routes, and sets research depth (shallow vs deep).
  • Shallow research agent — bounded, speed-optimized researcher with tool calling and citations.
  • Deep research agent — multi-phase workflow: advisory source routing, structured planning, concurrent researcher workers, bounded source-tool batching, and a dedicated writer.
  • Clarifier agent — report follow-up: questions about a finished report, rewrites, and delta research runs that carry the parent report's context.

Model wiring (exact ids) [aiq-github]: primary LLM for intent/routing/research and the default for planning/writing is nvidia/nemotron-3-super-120b-a12b; optional summarizer nvidia/nemotron-mini-4b-instruct; embeddings nvidia/llama-nemotron-embed-vl-1b-v2; optional vision-language model nvidia/nemotron-nano-12b-v2-vl. Alternative config profiles run GPT-OSS-120B or GPT-5.2 — the orchestration is model-agnostic at the config layer.

Pluggable RAG backends [aiq-github]: LlamaIndex + ChromaDB (the default web profile); "Foundational RAG" — i.e. the RAG blueprint's separately deployed query/ingestion servers; and OpenSearch built-in knowledge retrieval (self-hosted, Elasticsearch, or Amazon OpenSearch Service auth).

How you'd use it

  1. Clone the repo; run ./scripts/setup.sh (or manual uv pip install).
  2. Keys: NVIDIA_API_KEY required; Tavily/Serper/SerpAPI/SearchAPI optional for web search. Copy deploy/.env.example to deploy/.env and populate.
  3. Pick a YAML profile from configs/.
  4. Run it: ./scripts/start_cli.sh for the CLI, or ./scripts/start_e2e.sh for the web UI at localhost:3000 with the backend API at localhost:8000; Docker Compose (deploy/compose/) and Helm (deploy/helm/deployment-k8s) for services; a REST API runs deep-research jobs asynchronously; a 3-part notebook series lives in docs/notebooks/ [aiq-github].

What it demonstrates (why it's the one to study)

Citation-backed report generation with source tracking; report follow-up workflows (Q&A over a finished report, rewrites, delta research with parent context); evaluation against DeepResearch Bench and FreshQA; opt-in durable artifact capture (manifest-based file checkpointing to SQL or S3-compatible storage); and NAT-exported traces that preserve the agent hierarchy with model/tool usage for cost and latency analysis [aiq-github]. That last pair — benchmarked quality plus per-agent traces — is what separates a reference architecture from a demo.

When to reach for it vs building yourself

Reach for it when the product is research reports over enterprise corpora — it has already solved planning, worker fan-out, citation management, follow-up, and observability, and it plugs into a RAG backend you may already run. Build yourself when you need a simple agentic-RAG loop (grade-then-branch over one store — see the brief §3.2's LangGraph shape; AI-Q's planner/worker/writer machinery would be overhead), or when your orchestration standard isn't LangGraph/NAT. Its config-profile seam (swap Nemotron for GPT-class models) makes it a fair harness for comparing generators too.

Interview angle

"AI-Q is NVIDIA's deep-research blueprint: an orchestration node classifies intent and sets depth, then either a bounded shallow researcher or a deep agent with planning, concurrent researcher workers, and a dedicated writer — built on NeMo Agent Toolkit 1.8 and LangChain Deep Agents over LangGraph, generating with Nemotron 3 Super. What makes it credible as a reference is the operational rim: DeepResearch Bench / FreshQA evals, citation tracking, and NAT traces of the whole agent hierarchy for cost analysis. It's the agentic layer that treats the RAG blueprint as one pluggable backend."

Sources

aiq-github: {title: "AI-Q NVIDIA Blueprint — NVIDIA-AI-Blueprints/aiq README (v2.1.0; agents, model ids, backends, deploy scripts, evals)", url: "https://github.com/NVIDIA-AI-Blueprints/aiq", accessed: "2026-07-14"}
nat-github: {title: "NVIDIA NeMo Agent Toolkit — NVIDIA/NeMo-Agent-Toolkit (v1.8; nvidia-nat; profiling, MCP client/server)", url: "https://github.com/NVIDIA/NeMo-Agent-Toolkit", accessed: "2026-07-14"}