Qwen3-Embedding-0.6B-ONNX

Qwen3-Embedding-0.6B

This model is an ONNX conversion of the original Qwen3-Embedding model tailored for Sinequa usage.

Usage

This model is an Instruct Retriever, it's LLM Based. So you can give instruction at query time that better suits your needs. Only the query prefix should be added.

Packaging

Download this folder and use Package from local in the Sinequa Runnable Model Wizard.

Note: No MRL Cutoff is specified, so the embedding dimension out of this vectorizer will be 1024.

Inference Times

GPU Memory Usage

Requirements

• Minimal Sinequa version: 11.12

Example of a payload in the Runnable Model API

{
  "inputs": [
    {
      "text": "What is the capital of China?"
    },
    {
        "text": "Explain Gravity"
    }
  ],
  "options": {
    "context": "query",
    "passagePrefix": "",
    "queryPrefix": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:"
  }
}

Notice the instruction stating the model is given a "web search query". You could modify it to make it more domain-specific.

Maximum Sequence Length

In this repository, the maximum sequence length of the model has been set to 200 tokens, as Sinequa default. You could extend it up to 1024 following these simple steps:

1. update the maximum sequence length to 1024 in the sinequa.conf.json file
2. In the tokenizer.json file, update the property max_length, under truncation to 1024.