Class Embedder

Namespace: LMKit.Embeddings

Assembly: LM-Kit.NET.dll

A class designed for generating embeddings from text and image. It facilitates execution of tasks related to natural language and coding, such as semantic search, clustering, topic modeling, and classification. Supports embeddings from:

Plain text strings
Tokenized text
File attachments (e.g. images like PNG, JPEG, TIFF, and documents like TXT, HTML) when the underlying model provides image-embedding capabilities.

public sealed class Embedder : IEmbedder, IQueryEmbedder, IDocumentEmbedder, IImageEmbedder, ITokenizingEmbedder

Inheritance: object

Embedder

Implements: IEmbedder

IQueryEmbedder

IDocumentEmbedder

IImageEmbedder

ITokenizingEmbedder

Inherited Members: object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Examples

Example: Generate text embeddings

using LMKit.Model;
using LMKit.Embeddings;
using System;
// Load an embedding model (e.g. "embeddinggemma-300m", "harrier-oss:0.6b", "qwen3-embedding:0.6b")
LM model = LM.LoadFromModelID("embeddinggemma-300m");
// Create the embedder
Embedder embedder = new Embedder(model);
// Generate embedding for a single text
float[] embedding = embedder.GetEmbeddings("Machine learning is fascinating.");
Console.WriteLine($"Embedding dimension: {embedding.Length}");
Console.WriteLine($"First 5 values: [{string.Join(", ", embedding.Take(5).Select(v => v.ToString("F4")))}...]");

Example: Batch embeddings for similarity comparison

using LMKit.Model;
using LMKit.Embeddings;
using System;
using System.Collections.Generic;
LM model = LM.LoadFromModelID("embeddinggemma-300m");
Embedder embedder = new Embedder(model);
// Generate embeddings for multiple texts
var texts = new List<string>
{
"The cat sat on the mat.",
"A feline rested on the rug.",
"The stock market closed higher today."
};
float[][] embeddings = embedder.GetEmbeddings(texts);
// Calculate cosine similarity between first two texts
float similarity = VectorOperations.CosineSimilarity(embeddings[0], embeddings[1]);
Console.WriteLine($"Similarity between text 1 and 2: {similarity:F4}");
// Compare with unrelated text
float dissimilarity = VectorOperations.CosineSimilarity(embeddings[0], embeddings[2]);
Console.WriteLine($"Similarity between text 1 and 3: {dissimilarity:F4}");

Example: Image embeddings (multimodal model)

using LMKit.Model;
using LMKit.Embeddings;
using LMKit.Media.Image;
using System;
// Load a vision-enabled embedding model
LM model = LM.LoadFromModelID("nomic-embed-vision");
Embedder embedder = new Embedder(model);
// Generate embedding from an image
ImageBuffer image = ImageBuffer.Load("photo.jpg");
float[] imageEmbedding = embedder.GetEmbeddings(image);
Console.WriteLine($"Image embedding dimension: {imageEmbedding.Length}");

Remarks

Key Features

Generate embeddings from text strings via GetEmbeddings(string, CancellationToken)
Batch processing for multiple texts via GetEmbeddings(IEnumerable<string>, CancellationToken)
Image embeddings (when model supports) via GetEmbeddings(ImageBuffer, CancellationToken)
File attachment embeddings via GetEmbeddings(Attachment, CancellationToken)
Token-based embeddings for pre-tokenized input
Async variants for all methods

Common Use Cases

Semantic search: Find similar documents by comparing embedding vectors
Clustering: Group similar texts or images together
Classification: Use embeddings as features for ML classifiers
RAG systems: Generate embeddings for retrieval-augmented generation

The embedding dimension is determined by the model and can be queried via EmbeddingSize.

Constructors

Embedder(LM): Initializes a new instance of the Embedder class.

Properties

EmbeddingSize: Gets the dimension of the vectors this embedder produces.

Model: Gets the LM instance associated with this object.

ModelId: Gets a stable identifier for the underlying model or deployment.

SupportsImageEmbeddings: Gets a value indicating whether the underlying model can currently produce image embeddings.

SupportsTextEmbeddings: Gets a value indicating whether this embedder can embed text. Text embedding requires the model's text tokenizer; image-only embedding checkpoints carry none, so their text methods throw InvalidModelException. Check this before routing text to an embedder that may be image-only.

Methods

GetCosineSimilarity(IList<float>, IList<float>): Calculates the cosine similarity between two embedding vectors, representing the cosine of the angle between them in a multidimensional space.

GetDocumentEmbeddingsAsync(IEnumerable<string>, CancellationToken): Asynchronously generates embedding vectors for a batch of documents/passages, applying the model's document instruction prefix when it declares one.

GetDocumentEmbeddingsAsync(string, CancellationToken): Asynchronously generates the embedding vector for one document/passage, applying the model's document instruction prefix when it declares one. For models without a document-side prefix this is identical to GetEmbeddingsAsync(string, CancellationToken).

GetEmbeddings(Attachment, CancellationToken): Generates the embedding vector for a given file attachment. If the associated LM supports image embeddings, image attachments (for example, PNG, JPEG, TIFF) will be processed into embeddings; otherwise, text‑based attachments (for example, TXT, HTML) will be used.

GetEmbeddings(ImageBuffer, CancellationToken): Generates the embedding vector for a given image.

GetEmbeddings(IEnumerable<ImageBuffer>, CancellationToken): Generates embedding vectors for a collection of images.

GetEmbeddings(IEnumerable<IList<int>>, CancellationToken): Generates embedding vectors for a collection of tokenized texts. Each vector represents a text in a high-dimensional space, enabling various natural language processing tasks by capturing semantic meaning.

GetEmbeddings(IEnumerable<string>, CancellationToken): Generates embedding vectors for a collection of text strings. Each vector represents a text in a high-dimensional space, enabling various natural language processing tasks by capturing semantic meaning.

GetEmbeddings(IList<int>, CancellationToken): Generates the embedding vector for a given tokenized text. This vector represents the text in a high-dimensional space, enabling various natural language processing tasks by capturing semantic meaning.

GetEmbeddings(string, CancellationToken): Generates the embedding vector for a given text string. This vector represents the text in a high-dimensional space, enabling various natural language processing tasks by capturing semantic meaning.

GetEmbeddingsAsync(Attachment, CancellationToken): Asynchronously generates the embedding vector for a given file attachment. If the associated LM supports image embeddings, image attachments (for example, PNG, JPEG, TIFF) will be processed into embeddings; otherwise, text‑based attachments (for example, TXT, HTML) will be used.

GetEmbeddingsAsync(ImageBuffer, CancellationToken): Asynchronously generates the embedding vector for a given image.

GetEmbeddingsAsync(IEnumerable<ImageBuffer>, CancellationToken): Asynchronously generates embedding vectors for a collection of images.

GetEmbeddingsAsync(IEnumerable<IList<int>>, CancellationToken): Asynchronously generates embedding vectors for a collection of tokenized texts. Each vector represents a text in a high-dimensional space, enabling various natural language processing tasks by capturing semantic meaning.

GetEmbeddingsAsync(IEnumerable<string>, CancellationToken): Asynchronously generates embedding vectors for a collection of text strings. Each vector represents a text in a high-dimensional space, enabling various natural language processing tasks by capturing semantic meaning.

GetEmbeddingsAsync(IList<int>, CancellationToken): Asynchronously generates the embedding vector for a given tokenized text. This vector represents the text in a high-dimensional space, enabling various natural language processing tasks by capturing semantic meaning.

GetEmbeddingsAsync(string, CancellationToken): Asynchronously generates the embedding vector for a given text string. This vector represents the text in a high-dimensional space, enabling various natural language processing tasks by capturing semantic meaning.

GetQueryEmbeddings(string, CancellationToken): Generates the embedding vector for a query string, applying the model's query instruction prefix when available. Use this for search/retrieval queries to improve recall with models that support asymmetric query/passage embedding (e.g., Qwen3 Embedding, Nomic Embed, BGE v1.5). For passage/document embedding, use GetEmbeddings(string, CancellationToken) instead.

GetQueryEmbeddingsAsync(string, CancellationToken): Asynchronously generates the embedding vector for a query string, applying the model's query instruction prefix when available. Use this for search/retrieval queries to improve recall with models that support asymmetric query/passage embedding (e.g., Qwen3 Embedding, Nomic Embed, BGE v1.5). For passage/document embedding, use GetEmbeddingsAsync(string, CancellationToken) instead.

TokenizeForEmbedding(string, EmbeddingRole): Tokenizes a passage/document string exactly as GetEmbeddingsAsync(string, CancellationToken) does internally, returning the token sequence that the embedding model consumes.