NVIDIA-AI-Blueprints/deep-researcher-agent

The NVIDIA Deep Researcher Agent Blueprint 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.

What it solves

Deep Researcher Agent is a deployable research backend designed for governed research workflows. It solves the problem of generating high-quality, citation-backed research reports and quick answers from a variety of data sources while allowing teams to self-host the application boundary and maintain control over their models, data, and security policies.

How it works

Built on the NVIDIA NeMo Agent Toolkit and LangChain Deep Agents, the system uses a LangGraph-based state machine to orchestrate different research depths:

  • Orchestration Node: Classifies the user's intent and determines if the request requires a simple response (meta) or research (shallow vs. deep).
  • Shallow Research: A fast, tool-augmented agent that provides bounded, cited answers.
  • Structured Deep Research: A multi-phase process involving advisory source routing, structured planning, concurrent researcher workers, and a dedicated writer to produce comprehensive reports.
  • Report Follow-up: Allows users to ask questions about a completed report or perform "delta research" using the parent report as context.

Who it’s for

  • AI researchers and developers: Those building or extending agentic research workflows.
  • Enterprise teams: Organizations requiring tool-augmented research with strict citation requirements and governed workflows.
  • NeMo Agent Toolkit users: Developers exploring advanced multi-agent patterns.

Highlights

  • Modular Workflows: Agents (orchestration, shallow, deep, clarifier) are composable and can run standalone or as a pipeline.
  • Diverse Data Sources: Supports web search (Tavily, Exa, Nimble, You.com), academic paper search (Serper, SerpAPI, SearchAPI), and knowledge layers (OpenSearch, Azure AI Search, LlamaIndex).
  • Durable Outputs: Features provider-neutral sandbox execution for code and opt-in artifact capture to store generated files in SQL or S3-compatible storage.
  • Production-Ready: Includes REST endpoints, OAuth-protected MCP sources, NeMo Guardrails middleware for policy control, and OTEL exports for observability.
  • Evaluation Harnesses: Built-in benchmarks (e.g., FreshQA, DeepResearch) to measure and iterate on agent quality.

Related

  • Project
    yarin-zhang/AI-Gist

    A privacy-first, cross-platform prompt management tool that allows users to organize, optimize, and use dynamic AI prompts with local storage.

  • Project
    hashicorp/agent-skills

    A collection of specialized AI agent plugins called Skills that extend the capabilities of AI assistants to better support HashiCorp Terraform and Packer.

  • Project
    Robinson-AI-Lab/FridayOS1.0

    An AI-powered personal knowledge management system that uses natural language to automatically organize and retrieve information within a structured Obsidian vault.

  • Project
    ChatbotXIO/ChatbotX

    An open-source, omnichannel chat marketing platform that lets users build AI-powered agents and automated conversation flows for platforms like WhatsApp, Instagram, and Telegram.