Class EmbeddingPreset

Namespace

Qavren.Edge.Embeddings.Onnx

Assembly

Qavren.Edge.Embeddings.Onnx.dll

Everything the ONNX graph does not tell you. A wrong preset is a silent quality bug, so every field that changes the numbers is required.

public sealed record EmbeddingPreset : IEquatable<EmbeddingPreset>

Inheritance

object

EmbeddingPreset

Implements

IEquatable<EmbeddingPreset>

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Properties

AttentionMaskName

The graph's attention-mask input name. A binding key.

Dimensions

The embedding width. Fixed: this library implements no Matryoshka truncation.

DocumentPrefix

The instruction prefix the document-side generator prepends, or null.

Id

The stable id ById(string) resolves.

InputIdsName

The graph's token-id input name. A BINDING KEY: inputs are bound by name, never by position.

LayerNormEpsilon

Epsilon for PostPoolLayerNorm. 1e-5 is PyTorch's default and what nomic's reference uses. Read ONLY when PostPoolLayerNorm is set, so three of the four shipped presets never touch it.

LowerCase

Whether the model was trained on lower-cased text.

Manifest

The pinned manifest: repo, commit SHA, per-file size and digest, SPDX licence.

MaxSequenceLength

The longest sequence the graph accepts, special tokens included.

ModelFile

The manifest-relative path of the graph ORT is handed.

Normalize

L2-normalise the pooled vector. True for all four shipped presets.

OutputName

The graph's output name. A binding key.

Pooling

How the batch pools. From the preset, NEVER a default: bge-small is CLS and everything else here is mean, and getting it wrong is a silent retrieval regression with no exception anywhere.

PostPoolLayerNorm

Apply LayerNorm(Span<float>, float) between pooling and L2 normalisation. Not Matryoshka truncation, which is cut - this is an unconditional step in nomic-embed-text-v1.5's reference pooling. Its modules.json carries only Transformer

• Pooling (no Normalize module) and its documented pipeline is mean-pool, then F.layer_norm over the 768 dim, then L2. Omitting it ships plausible-but-off vectors with no error anywhere, which is exactly what the required fields on this record exist to prevent. False for the three MiniLM/BGE presets.

QueryPrefix

The instruction prefix the query-side generator prepends, or null.

SequenceBuckets

The padded widths a batch rounds up to, ascending. What the batch assembler rounds to and what PinnedSequenceLength is validated against.

TokenTypeIdsName

The graph's token-type input name, or null when the graph declares no such input - in which case the generator binds two tensors instead of three. Bound BY NAME, never by position: nomic declares (input_ids, token_type_ids, attention_mask) while MiniLM declares (input_ids, attention_mask, token_type_ids).

TokenizerFile

The manifest-relative path of the tokenizer asset.

TokenizerKind

Which tokenizer family TokenizerFile is.