Class OnnxEmbeddingOptions

Namespace

Qavren.Edge.Embeddings.Onnx

Assembly

Qavren.Edge.Embeddings.Onnx.dll

Everything AddOnnxEmbeddings configures. Every default here changes behaviour.

public sealed class OnnxEmbeddingOptions

Inheritance

object

OnnxEmbeddingOptions

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Properties

DefaultInputKind

Which prefix the unkeyed generator applies. The query sibling is reached through QueryServiceKey.

MaxBatchSize

How many inputs one ORT Run carries. 16 by default.

MaxConcurrency

Concurrent ORT Runs. 1 by default: ORT already parallelises intra-op, and a second inference doubles peak native memory on a phone.

ModelSource

Null uses the HTTP model source over the preset's manifest.

PinnedSequenceLength

Null (the default) uses SequenceBuckets: the graph keeps its declared symbolic sequence_length, CoreML partitions normally, and each batch is padded to the smallest bucket that fits it.

Set to one of the buckets to pin the shape instead. That emits AddFreeDimensionOverrideByName("sequence_length", N) and ("batch_size", MaxBatchSize) into FreeDimensionOverrides and turns CoreMlProviderOptions.RequireStaticInputShapes on. One pinned shape means one session and one CoreML compile, and every batch - including a batch of one short string - pays a full N-token run. It is a measured trade, not a default.

Preset

The preset. Defaults to MiniLmL6V2Int8.

Session

The per-model session settings this preset's model is registered with.

ShrinkBatchUnderMemoryPressure

Halve the effective batch size after a latched memory pressure until it clears.

Truncation

What happens to an over-long input. Truncate-and-log by default.