Gst-nvdstextembedder

The Gst-nvdstextembedder plugin is a DeepStream-compatible GStreamer Python plugin that generates text embeddings, designed to work alongside Gst-nvdsvisionencoder-c for text-to-image search applications. It supports text-to-image search, person-attribute search (e.g. “person wearing red shirt”), zero-shot classification, and semantic search. Texts can be set via the texts property, or encoded on demand using a custom GStreamer “embedding” query that returns results synchronously without modifying element properties.

Inputs and Outputs

Inputs
- texts property, or a custom “embedding” GstQuery, containing a JSON array of text strings to encode
Control parameters
- device, model-name, batch-size, use-fp16, verbose
Output
- embeddings property containing a JSON array of computed embedding vectors (or the synchronous query result for on-demand encoding)

Features

The following table summarizes the features of the plugin.

Gst-nvdstextembedder plugin features

Feature	Description	Release
Text-to-embedding encoding	Generates text embedding vectors using SigLIP2 Giant (1536-dim, default) or DFN5B-CLIP (1024-dim) models	DS 9.2
On-demand embedding query	Supports a custom GstQuery for synchronous on-demand text encoding without changing element properties	DS 9.2
Vision-encoder pairing	Designed to work alongside Gst-nvdsvisionencoder-c so text and image embeddings share a common vector space	DS 9.2

Gst Properties

The following table describes the Gst properties of the Gst-nvdstextembedder plugin.

Gst-nvdstextembedder plugin gst properties

Property	Meaning	Type and Range	Example Notes
texts	JSON array of texts to encode (e.g. ’[“query1”, “query2”]‘)	String	texts=’[“person wearing red shirt”]‘
embeddings (read-only)	JSON array of computed embeddings	String
batch-size	Maximum batch size for inference	Integer, 1 to 128	batch-size=64
verbose	Enables verbose logging	Boolean	verbose=0
device	Device used for inference (cuda or cpu)	String	device=cuda
use-fp16	Enables FP16 inference for faster processing	Boolean	use-fp16=1
model-name	Encoder selection, set from curl/API (e.g. an ONNX model name). Use a name containing “siglip2-giant” or “giant” for SigLIP2 Giant (1536-dim), or “dfn5b-clip” for DFN5B-CLIP (1024-dim)	String	model-name=siglip2-giant
embedding-dim (read-only)	Model embedding dimension	Integer, 0 to 65536
num-texts (read-only)	Number of texts currently loaded	Integer, 0 to 2,147,483,647
model-loaded (read-only)	Whether the model is loaded and ready	Boolean
last-query-result (read-only)	JSON string of the last embedding query result	String
onnx-model-path	Path to the DFN-CLIP ONNX model file (used when model-name contains “dfnclip”)	String	onnx-model-path=/path/to/model.onnx
tokenizer-dir	Path to the tokenizer directory with tokenizer.json (used when model-name contains “dfnclip”)	String	tokenizer-dir=/path/to/tokenizer