nomic-embed-text-v1.5: Resizable Production Embeddings with Matryoshka Representation Learning

nomic-embed-text-v1.5 is an improvement upon Nomic Embed that utilizes Matryoshka Representation Learning which gives developers the flexibility to trade off the embedding size for a negligible reduction in performance.

Name	SeqLen	Dimension	MTEB
nomic-embed-text-v1	8192	768	62.39
nomic-embed-text-v1.5	8192	768	62.28
nomic-embed-text-v1.5	8192	512	61.96
nomic-embed-text-v1.5	8192	256	61.04
nomic-embed-text-v1.5	8192	128	59.34
nomic-embed-text-v1.5	8192	64	56.10

Hosted Inference API

The easiest way to get started with Nomic Embed is through the Nomic Embedding API.

Generating embeddings with the nomic Python client is as easy as

from nomic import embed

output = embed.text(
    texts=['Nomic Embedding API', '#keepAIOpen'],
    model='nomic-embed-text-v1.5',
    task_type='search_document',
    dimensionality=256,
)

print(output)

For more information, see the API reference

Data Visualization

Click the Nomic Atlas map below to visualize a 5M sample of our contrastive pretraining data!

Training Details

We train our embedder using a multi-stage training pipeline. Starting from a long-context BERT model, the first unsupervised contrastive stage trains on a dataset generated from weakly related text pairs, such as question-answer pairs from forums like StackExchange and Quora, title-body pairs from Amazon reviews, and summarizations from news articles.

In the second finetuning stage, higher quality labeled datasets such as search queries and answers from web searches are leveraged. Data curation and hard-example mining is crucial in this stage.

For more details, see the Nomic Embed Technical Report and corresponding blog post.

Training data to train the models is released in its entirety. For more details, see the contrastors repository

Usage

Note nomic-embed-text requires prefixes! We support the prefixes [search_query, search_document, classification, clustering]. For retrieval applications, you should prepend search_document for all your documents and search_query for your queries.

Sentence Transformers

import torch.nn.functional as F
from sentence_transformers import SentenceTransformer

matryoshka_dim = 512

model = SentenceTransformer("nomic-ai/nomic-embed-text-v1.5", trust_remote_code=True)
sentences = ['search_query: What is TSNE?', 'search_query: Who is Laurens van der Maaten?']
embeddings = model.encode(sentences, convert_to_tensor=True)
embeddings = F.layer_norm(embeddings, normalized_shape=(embeddings.shape[1],))
embeddings = embeddings[:, :matryoshka_dim]
embeddings = F.normalize(embeddings, p=2, dim=1)
print(embeddings)

Transformers

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel

def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0]
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

sentences = ['search_query: What is TSNE?', 'search_query: Who is Laurens van der Maaten?']

tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
model = AutoModel.from_pretrained('nomic-ai/nomic-embed-text-v1.5', trust_remote_code=True, safe_serialization=True)
model.eval()

encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

+ matryoshka_dim = 512

with torch.no_grad():
    model_output = model(**encoded_input)

embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
+ embeddings = F.layer_norm(embeddings, normalized_shape=(embeddings.shape[1],))
+ embeddings = embeddings[:, :matryoshka_dim]
embeddings = F.normalize(embeddings, p=2, dim=1)
print(embeddings)

The model natively supports scaling of the sequence length past 2048 tokens. To do so,

- tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
+ tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased', model_max_length=8192)


- model = AutoModel.from_pretrained('nomic-ai/nomic-embed-text-v1', trust_remote_code=True)
+ model = AutoModel.from_pretrained('nomic-ai/nomic-embed-text-v1', trust_remote_code=True, rotary_scaling_factor=2)

Transformers.js

import { pipeline, layer_norm } from '@xenova/transformers';

// Create a feature extraction pipeline
const extractor = await pipeline('feature-extraction', 'nomic-ai/nomic-embed-text-v1.5', {
    quantized: false, // Comment out this line to use the quantized version
});

// Define sentences
const texts = ['search_query: What is TSNE?', 'search_query: Who is Laurens van der Maaten?'];

// Compute sentence embeddings
let embeddings = await extractor(texts, { pooling: 'mean' });
console.log(embeddings); // Tensor of shape [2, 768]

const matryoshka_dim = 512;
embeddings = layer_norm(embeddings, [embeddings.dims[1]])
    .slice(null, [0, matryoshka_dim])
    .normalize(2, -1);
console.log(embeddings.tolist());

Join the Nomic Community

• Nomic: https://nomic.ai
• Discord: https://discord.gg/myY5YDR8z8
• Twitter: https://twitter.com/nomic_ai

Citation

If you find the model, dataset, or training code useful, please cite our work

@misc{nussbaum2024nomic,
      title={Nomic Embed: Training a Reproducible Long Context Text Embedder}, 
      author={Zach Nussbaum and John X. Morris and Brandon Duderstadt and Andriy Mulyar},
      year={2024},
      eprint={2402.01613},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}