Table of Contents

What Embedding Models Does LM-Kit.NET Support and When Should I Use Them?


TL;DR

LM-Kit.NET supports text embeddings (7 models from 67 MB to 4.7 GB), image embeddings (1 model, 97 MB), and reranking (1 model, 438 MB). Use text embeddings for RAG, semantic search, and classification. Use image embeddings for visual similarity search. Use the reranker to refine initial search results for higher accuracy. For most applications, Gemma Embedding 300M or BGE-M3 offer the best balance of quality and size.


Text Embedding Models

Model ID Size Dimension Context Languages Best For
bge-small 68 MB 384 512 English Lightweight, CPU-friendly, low-latency
nomic-embed-text 90 MB 768 2,048 Multi Production embeddings with Matryoshka support
embeddinggemma-300m 303 MB 768 2,058 100+ Best all-round compact model, Matryoshka support
bge-m3 438 MB 1024 8,192 100+ Multilingual, long context, dense+sparse retrieval
qwen3-embedding:0.6b 639 MB Variable 32,768 Multi Long-context semantic search
qwen3-embedding:4b 2.5 GB Variable 40,960 Multi Mid-size, multilingual, long-context
qwen3-embedding:8b 4.7 GB Variable 40,960 Multi Highest quality, reranking, classification, clustering

Matryoshka Representation Learning

Models marked with Matryoshka support (Gemma Embedding, Nomic Embed) can truncate their output dimensions (e.g., 768 → 512 → 256 → 128) with minimal quality loss. This lets you trade embedding quality for storage size and search speed.


Image Embedding Model

Model ID Size Dimension Architecture Best For
nomic-embed-vision 97 MB 768 ViT-B/16 Visual similarity search, multimodal RAG

Nomic Embed Vision is aligned with Nomic Embed Text, enabling cross-modal search: find images by text description or find text by image similarity.

using LMKit.Model;
using LMKit.Embeddings;

using LM visionModel = LM.LoadFromModelID("nomic-embed-vision");
var embedder = new Embedder(visionModel);

// Get image embedding
float[] imageVector = embedder.GetEmbeddings(new ImageBuffer("product-photo.jpg"));

// Compare with text embedding (using nomic-embed-text)
float[] textVector = textEmbedder.GetEmbeddings("red leather handbag");

float similarity = VectorOperations.CosineSimilarity(imageVector, textVector);

Reranking Model

Model ID Size Context Languages Best For
bge-m3-reranker 438 MB 8,192 100+ Refining RAG search results for higher precision

The reranker scores query-document pairs with a cross-encoder, producing more accurate relevance scores than initial embedding similarity:

using LMKit.Embeddings;

using LM rerankerModel = LM.LoadFromModelID("bge-m3-reranker");
var reranker = new Reranker(rerankerModel);

// Score a single document against a query
float score = reranker.GetScore("What is the return policy?", "Our return policy allows...");

// Rerank a collection with blending (alpha controls mix of original vs reranker scores)
var reranked = reranker.Rerank(query, searchResults, alpha: 0.5f);

Which Model Should I Choose?

Use Case Recommended Model Why
English-only RAG, limited resources bge-small Smallest, fastest, CPU-friendly
General-purpose RAG embeddinggemma-300m Best quality for its size, 100+ languages
Multilingual RAG with long documents bge-m3 8K context, 100+ languages, dense+sparse
Maximum retrieval quality qwen3-embedding:8b Largest, highest accuracy
Image similarity search nomic-embed-vision Only image embedding model
Refine search results bge-m3-reranker Cross-encoder reranking

📚 Related Content

Share