CKG-NVIDIA- AI - Nvidia Developer Stack As A Compressed Knowledge Graph - 20 Domains, 998 Nodes, Agents Traverse Type Dependency Graphs
**NVIDIA AI Developer Stack as a Compressed Knowledge Graph ** 4× F1 · 11× fewer tokens · 998 nodes · deterministic traversal
Instead of scanning docs or re-inferring structure on every query, your agent traverses declared relationships. 20 NVIDIA AI domains, every prerequisite chain typed and ready.
Install: uvx ckg-nvidia-ai
Example: query_ckg("TensorRT-LLM", "nvidia-tensorrt-triton", depth=3)
Returns: CUDA Toolkit, FP8/FP4 Quantization, Hopper SM90 as hard prerequisites. Triton Inference Server and NIM Microservice Runtime as dependents. 269 tokens. No hallucination.
If an edge isn't declared, the traversal returns nothing rather than hallucinating a path. That silence is signal.
**Tools: **list_domains · query_ckg · get_prerequisites · search_concepts · ask_nvidia (Ollama, no API key)
Domains: NIM · NeMo · TensorRT · CUDA · Isaac · Cosmos · Omniverse · Riva · Jetson · DRIVE · Clara · Metropolis · GameWorks · HPC SDK · CUDA-X · Developer Tools · Graphics Research · AI Enterprise · Developer Ecosystem · OpenShell
Benchmark: CKG F1 0.471 vs RAG 0.123 · 269 tokens vs 2,982 per query
服务器配置
{
"mcpServers": {
"nvidia-ai": {
"command": "uvx",
"args": [
"ckg-nvidia-ai"
]
}
}
}