Skip to main content

​
Overview

The NVIDIA AI-Q Blueprint is an open reference implementation for enterprise research agents. It connects agents to internal data, reasons over it with NVIDIA-served models, and returns sourced business insight. It runs on the NeMo Agent Toolkit, so its tools are nat functions declared in a workflow YAML. Nimble ships as the nimble_web_search function, a built-in data-source plugin. Calls go to Nimble’s Search API and come back as <Document> blocks the agent can cite. No fork, no custom tool code. The provider ships in AI-Q v2.2.0 and later.

​
What the integration unlocks

Any AI-Q agent gains live web retrieval without tool-specific code. The model passes a natural-language question, and the provider returns ranked results as escaped XML:
Every block carries an absolute href, and the provider drops results without one. Nimble handles anti-bot evasion, JavaScript rendering, and geo-targeting underneath.

​
Quick Start

​
1. Get a Nimble API key

Get your API key from Nimble’s dashboard (free trial available) and export it in the environment AI-Q runs in:
Set this before starting AI-Q. Without it the tool still registers and logs one warning, then returns an error the first time the agent searches. Startup looks healthy while the agent cannot search, so treat that warning as a failure.

​
2. Install AI-Q with the provider

setup.sh creates the virtual environment and installs every data-source plugin, nimble_web_search included. To add it to a checkout you already have:

​
3. Add the function to your workflow

Declare nimble_web_search under functions in your workflow YAML:
Then reference web_search_tool from the agent that should search.

​
4. Confirm the tool registered

A row for nimble_web_search means the plugin entry point resolved.

​
Configuration

Optional in config, required in practice. Falls back to the NIMBLE_API_KEY environment variable. A key set here is passed straight to the SDK and never written to the process environment.
Optional. Results to return, 1 to 100. Defaults to 5. Sent unclamped, so Nimble validates the range and reports the error.
Optional. lite, fast, or deep. Defaults to lite.
  • lite: title, URL, and description only. Lowest latency and lowest token cost.
  • fast: rich content at low latency, and the depth Nimble recommends for agent loops.
  • deep: full page content per result. Returns the raw page, including navigation and image data, so pair it with max_content_length.
See Search Depth for the tradeoffs and Pricing for what each depth costs.
Optional. Defaults to general. Leave it there for research and general agent use.The provider accepts general, news, location, shopping, geo, and social, and validates the value at config load, so a typo surfaces at startup with the valid set named in the error.focus is a workflow setting, not an agent-chosen parameter. The model only supplies a query, so a general research question cannot silently reroute.
Optional. Characters kept per result, truncated with an ellipsis. Defaults to 10000. Set to None to disable truncation. Matters most with search_depth: deep.
Optional. Attempts on a transient failure, 1 to 10. Defaults to 3, with exponential backoff capped at 30 seconds. Authentication and authorization errors return immediately instead of retrying.
Optional. country is an ISO 3166 code for geo-targeted results, default US. locale is a language code, default en.
NVIDIA maintains the full parameter table in its own configuration reference.

​
Additional Resources

AI-Q Configuration Reference

NVIDIA’s own parameter table for nimble_web_search.

AI-Q on GitHub

The blueprint, the provider source, and its tests.

Nimble Search API

The endpoint behind the provider: parameters and response format.

Nimble MCP Server

Reach Extract, Map, and Crawl over the Model Context Protocol.