NVIDIA/game-agent-sdk

NVIDIA/game-agent-sdk

Description: NVIDIA Game Agent SDK

Language: C++

License: Apache-2.0

Stars: 0

Forks: 0

Open issues: 0

Created: 2026-05-26T17:19:25Z

Pushed: 2026-06-16T03:06:31Z

Default branch: main

Fork: no

Archived: no

README:

Game Agent SDK

The Game Agent SDK provides on-device building blocks for AI companions, with a C99 ABI and a thin C++ wrapper for natural integration into game engines and other native applications. A companion uses AI inferencing locally on the end user's GPU with a small language model (SLM). The SDK exposes three groups of APIs:

• Agent APIs: stateful APIs that own conversation history and drive the inference loop autonomously, including any tool-call chains, with minimal application code.
• Chat APIs: stateless APIs that give developers direct control over inference.
• RAG APIs: enable semantic, lexical, and hybrid knowledge retrieval from developer-built databases, grounding model responses in up-to-date knowledge.

Requirements

• Windows 10 or 11
• NVIDIA Ampere GPU or newer with sufficient VRAM (the bundled SLM uses ~3.8 GB)
• NVIDIA driver 570.65 or later
• Cuda 12.8 SDK or later

Quick start

Build the Project

Building from source requires Visual Studio 2019 or later (or the equivalent VS Build Tools).

setup.bat build.bat

setup.bat generates Visual Studio project files under _project/vs2022/. build.bat builds the Release configuration; pass -debug for Debug.

Get the Models

The four models bundled with the SDK (one SLM, two embedders, and a cross-encoder reranker; ~3 GB total) are not stored in this repository. Download the source zip from the NVIDIA Developer Portal and extract data/models/ into your checkout. The source zip is a complete working tree, so it can also be used directly without a separate git clone.

Run a Sample

Sample executables land in _out/. Agent-Sample is a good place to begin, a poker assistant that demonstrates the Agent APIs:

cd _out Agent-Sample.exe

The available samples demonstrate how the SDK can be integrated for various use cases; they are intended as starting points and further tuning would be needed to improve response quality in a real application.

• Agent-Sample.exe: poker assistant on the Agent APIs. The SDK drives the inference-and-tool-call loop across two tools: search (poker knowledge, backed by RAG) and check_status (game state).
• Chat-Sample.exe: a simplified version of the same poker assistant on the Chat APIs, with the application driving the loop directly. Same prompt, tools, and RAG pipeline.
• MultiAgent-Sample.exe: two Agents sharing one SLM, with distinct identities and conversation histories.

Run any sample with -h for command-line options. For a full walkthrough of all samples, see [Samples](docs/sphinx/source/samples.md).

BuildRag.exe is also included -- a tool for building a semantic or lexical RAG database from raw text files. See [Building a RAG database](docs/sphinx/source/rag-building-a-database.md).

Documentation

For API details, architecture, and in-depth sample walkthroughs, the documentation is the best place to start beyond this README:

• [Introduction](docs/sphinx/source/introduction.md)
• [API overview](docs/sphinx/source/api-overview.md)
• [Agent deep dive](docs/sphinx/source/agent-deep-dive.md)
• [Samples](docs/sphinx/source/samples.md)
• [Building a RAG database](docs/sphinx/source/rag-building-a-database.md)
• [Limitations](docs/sphinx/source/limitations.md)
• [Troubleshooting](docs/sphinx/source/troubleshooting.md)

Packaging a release

package.bat creates a redistributable SDK zip with the version number you supply when prompted. package_nomodels.bat creates a smaller zip without the bundled models.

License

The Game Agent SDK is licensed under the Apache License 2.0 (see [LICENSE.txt](LICENSE.txt)). Third-party components and their licenses are listed in [THIRD-PARTY-LICENSES.md](THIRD-PARTY-LICENSES.md).