Redis
redis-vl-dotnet Documentation redis-vl-dotnet

redis-vl-dotnet

ONNX Vectorizer

RedisVL.Vectorizers.Onnx provides OnnxTextVectorizer, an IBatchTextVectorizer implementation that generates sentence embeddings locally with ONNX Runtime. It lets applications embed text entirely offline, without an API key or network call, mirroring the SentenceTransformers workflow.

Package contents

  • OnnxTextVectorizer for local sentence embedding generation

  • OnnxVectorizerOptions for model asset and runtime configuration

  • OnnxPoolingStrategy for selecting mean or [CLS] pooling

  • OnnxRuntimeSessionOptions for ONNX Runtime session tuning

  • OnnxVectorizerPackage as the package marker type

Local model assets

This package does not bundle model files. Applications must provide:

  • a local model.onnx file

  • a local Hugging Face tokenizer.json file

A typical source is a SentenceTransformers model such as sentence-transformers/all-MiniLM-L6-v2 exported to ONNX (for example with Hugging Face Optimum), which produces both a model.onnx and a tokenizer.json.

The reference workflow assumes a BERT-style tokenizer and embedding shape:

  • [CLS] text [SEP]

  • input_ids

  • attention_mask

  • token_type_ids when the model expects them

  • a last_hidden_state (or token_embeddings) output of shape [batch, sequence, hidden]

Request options

OnnxVectorizerOptions supports:

  • ModelPath for the local ONNX model file

  • TokenizerPath for the local tokenizer definition

  • MaxSequenceLength with a default of 512

  • Pooling with a default of OnnxPoolingStrategy.Mean; choose OnnxPoolingStrategy.Cls for models that expose a [CLS]-pooled representation

  • Normalize with a default of true, which L2-normalizes each embedding so cosine similarity equals the dot product

  • SessionOptions for local ONNX Runtime tuning such as graph optimization, execution mode, and thread counts

Mean pooling averages the per-token outputs using the attention mask, matching the most common SentenceTransformers configuration.

Example workflow

/examples/OnnxVectorizerExample shows the expected integration pattern:

  • embed a batch of prompts in a single call

  • print the embedding count and dimensionality

  • compare cosine similarity between a paraphrase and an unrelated prompt

Run it from the repository root after setting local asset paths:

export ONNX_VECTORIZER_MODEL_PATH=/path/to/model.onnx
export ONNX_VECTORIZER_TOKENIZER_PATH=/path/to/tokenizer.json
dotnet run --project examples/OnnxVectorizerExample/OnnxVectorizerExample.csproj

Validation references

  • tests/RedisVL.Tests/Vectorizers/OnnxTextVectorizerTests.cs covers validation, pooling, normalization, and batch behavior with stubbed inference

  • tests/RedisVL.Tests/Vectorizers/OnnxTextTokenizerTests.cs covers tokenizer [CLS]/[SEP] wrapping and truncation

  • tests/RedisVL.Tests/Vectorizers/OnnxTextVectorizerSmokeTests.cs is the local-asset smoke test gate