Dhara.AI.Inference 0.2.0

Dhara.AI.Inference

Low-level local inference primitives for Dhara.AI packages.

Dhara.AI.Inference is the ONNX and tokenizer layer behind the higher-level embedding APIs. It is useful when you want direct control over model paths, tokenization, ONNX Runtime session options, pooling, and normalization without taking a dependency on the convenience APIs in Dhara.AI.LocalEmbeddings.

What It Provides

• OnnxTextEmbeddingModel for running BERT-style text embedding models from an ONNX file.
• OnnxTextEmbeddingOptions for model paths, vocabulary paths, input/output names, pooling, normalization, and ONNX session customization.
• BertEmbeddingTokenizer for vocabulary-based tokenization into input_ids, attention_mask, and token_type_ids.
• EmbeddingMath for cosine similarity, mean pooling, and vector normalization.
• EmbeddingPoolingMode for mean-pooling or CLS-token embedding extraction.

Basic Usage

using Dhara.AI.Inference;

var options = new OnnxTextEmbeddingOptions
{
    ModelPath = "LocalEmbeddingsModel/default/model.onnx",
    VocabularyPath = "LocalEmbeddingsModel/default/vocab.txt",
    ModelId = "default",
    PoolingMode = EmbeddingPoolingMode.Mean,
    NormalizeEmbeddings = true
};

using var model = new OnnxTextEmbeddingModel(options);
var vectors = model.Embed(
[
    "local embeddings for semantic search",
    "native AOT friendly inference in .NET"
]);

var score = EmbeddingMath.CosineSimilarity(vectors[0], vectors[1]);

Model Compatibility

The model is expected to accept BERT-style inputs such as input_ids and attention_mask. If the ONNX graph exposes token_type_ids, the runtime sends them automatically; if it does not, the runtime omits them. This keeps the runtime compatible with many sentence-transformer exports, including models that produce 384-dimensional or 768-dimensional vectors.

The runtime supports output tensors shaped as either:

• [batch, dimensions] when the model already returns sentence embeddings.
• [batch, sequence, dimensions] when the runtime should pool token embeddings.

Native AOT

The package is marked as AOT-compatible. Native AOT is validated by publishing an application that consumes the package, because AOT behavior is determined at the application boundary.

Related Package

Most application code should start with Dhara.AI.LocalEmbeddings, which wraps this runtime with Microsoft.Extensions.AI, model acquisition, compact storage formats, and ranking helpers.