deepstream-generate-pipeline

Describe your input source, inference needs, tracker, output, and platform in natural language. The skill retrieves matching verified pipelines with a BM25 script, assembles the gst-launch-1.0 command, and validates it before presenting a single-line runnable pipeline.

Request something like "detect and track on 4 RTSP streams with NvDCF". The generated pipeline uses uridecodebin sources feeding nvstreammux named pads, nvinfer for detection, nvtracker with the chosen tracker config, and nvmultistreamtiler for tiled output.

Yes, it targets x86 dGPU (T4, A100, L40, RTX) and aarch64 platforms including Jetson Orin/Xavier/Nano and SBSA servers like Grace and GH200. Platform selection determines the display sink: nveglglessink for dGPU and nv3dsink for Jetson/SBSA.

The scripts use only the Python 3.8+ standard library with no pip packages. DeepStream SDK must be installed at /opt/nvidia/deepstream/deepstream/ with gst-launch-1.0 and gst-inspect-1.0 on PATH for element verification and live parse validation.

Multi-stream pipelines use named pad references like m.sink_0 on nvstreammux, and fakesrc cannot negotiate caps through named pads, which would cause false-positive linking errors. Syntax, element, and property checks still run for these pipelines.

The retriever uses no embeddings or semantic model, so unusual queries may return low confidence. In that case the skill falls back to documented assembly rules covering sources, inference chains, trackers, sinks, and extra operations instead of retrieved examples.