snowflake-arctic-embed-l

1. Multilingual without compromise: Excels in English and non-English retrieval, outperforming leading open-source and proprietary models on benchmarks like MTEB Retrieval, CLEF, and MIRACL.

2. Inference efficiency: Its 303m non-embedding parameters inference is fast and efficient for any scale.

3. Compression-friendly: Achieves high-quality retrieval with embeddings as small as 128 bytes/vector using Matryoshka Representation Learning (MRL) and quantization-aware embedding training. Please note that like our v1.5 model, the MRL for this model is 256 dimensions, and high-quality 128-byte compression is achieved via 4-bit quantization (e.g. using a pq256x4fs fast-scan FAISS index or using the example code published alongside our 1.5 model).

4. Drop-In Replacement: arctic-embed-l-v2.0 builds on BAAI/bge-m3-retromae which allows direct drop-in inference replacement with any form of new libraries, kernels, inference engines etc.

5. Long Context Support: arctic-embed-l-v2.0 builds on BAAI/bge-m3-retromae which can support a context window of up to 8192 via the use of RoPE.

Quality Benchmarks

Unlike most other open-source models, Arctic-embed-l-v2.0 excels across English (via MTEB Retrieval) and multilingual (via MIRACL and CLEF). You no longer need to support models to empower high-quality English and multilingual retrieval. All numbers mentioned below are the average NDCG@10 across the dataset being discussed.

Aside from high-quality retrieval arctic delivers embeddings that are easily compressible. Leverage vector truncation via MRL to decrease vector size by 4x with less than 3% degredation in quality. Combine MRLed vectors with vector compression (Int4) to power retrieval in 128 bytes per doc.

Usage

Using Sentence Transformers

from sentence_transformers import SentenceTransformer

# Load the model
model_name = 'Snowflake/snowflake-arctic-embed-l-v2.0'
model = SentenceTransformer(model_name)

# Define the queries and documents
queries = ['what is snowflake?', 'Where can I get the best tacos?']
documents = ['The Data Cloud!', 'Mexico City of Course!']

# Compute embeddings: use `prompt_name="query"` to encode queries!
query_embeddings = model.encode(queries, prompt_name="query")
document_embeddings = model.encode(documents)

# Compute cosine similarity scores
scores = model.similarity(query_embeddings, document_embeddings)

# Output the results
for query, query_scores in zip(queries, scores):
    doc_score_pairs = list(zip(documents, query_scores))
    doc_score_pairs = sorted(doc_score_pairs, key=lambda x: x[1], reverse=True)
    print("Query:", query)
    for document, score in doc_score_pairs:
        print(score, document)

Using Huggingface Transformers

You can use the transformers package to use Snowflake's arctic-embed model, as shown below. For optimal retrieval quality, use the CLS token to embed each text portion and use the query prefix below (just on the query).

import torch
from transformers import AutoModel, AutoTokenizer

model_name = 'Snowflake/snowflake-arctic-embed-l-v2.0'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name, add_pooling_layer=False)
model.eval()

query_prefix = 'query: '
queries  = ['what is snowflake?', 'Where can I get the best tacos?']
queries_with_prefix = ["{}{}".format(query_prefix, i) for i in queries]
query_tokens = tokenizer(queries_with_prefix, padding=True, truncation=True, return_tensors='pt', max_length=8192)

documents = ['The Data Cloud!', 'Mexico City of Course!']
document_tokens =  tokenizer(documents, padding=True, truncation=True, return_tensors='pt', max_length=8192)

# Compute token embeddings
with torch.no_grad():
    query_embeddings = model(**query_tokens)[