Hugging Face TEI

¶Sparse Embeddings with Text Embedding Inference Client

The TeiClient provides a method called embedSparse() to generate sparse embeddings for input text. Sparse embeddings are optimized for use cases like information retrieval and search indexing, where sparsity helps with efficient storage and fast similarity computations.

Tip

Sparse embeddings contain many zero values and only a few non-zero components, making them memory-efficient and ideal for large-scale retrieval systems.

¶Example: Generate Sparse Embeddings

use Partitech\PhpMistral\Clients\Tei\TeiClient;
use Partitech\PhpMistral\MistralClientException;

// Load TEI SPLADE pooling endpoint and API key
$teiUrl = getenv('TEI_SPLADE_POOLING_URI');
$apiKey = getenv('TEI_API_KEY');

// Instantiate the TEI client
$client = new TeiClient(apiKey: (string) $apiKey, url: $teiUrl);

try {
    $sparseEmbedding = $client->embedSparse(inputs: 'What is Deep Learning?');
    var_dump($sparseEmbedding);
} catch (MistralClientException $e) {
    echo $e->getMessage();
}

¶Output

array(1) {
  [0]=>
  array(1190) {
    [0]=>
    array(2) {
      ["index"]=>
      int(1003)
      ["value"]=>
      float(0.2055648)
    }
    [1]=>
    array(2) {
      ["index"]=>
      int(1005)
      ["value"]=>
      float(0.34785014)
    }
    [2]=>
    array(2) {
      ["index"]=>
      int(1006)
      ["value"]=>
      float(0.6190006)
    }
    ...
  }
}

• Each embedding consists of a list of (index, value) pairs.
• index: Refers to the position of a token or feature in the vocabulary.
• value: Represents the strength or weight of that feature.

Note

Unlike dense embeddings (e.g., 768 or 1024 dimensions with all values filled), sparse embeddings only contain non-zero values, significantly reducing memory usage.

¶Use Cases

• Search Engines: Efficiently match queries against large document corpora using sparse vector search.
• Information Retrieval: Build scalable retrieval systems with high-performance sparse embeddings.
• Hybrid Search: Combine sparse and dense embeddings to leverage the benefits of both approaches (BM25-like sparse scoring + semantic dense scoring).

¶Differences Between Dense and Sparse Embeddings

Important

Make sure your downstream system (e.g., vector database, search engine) supports sparse vectors if you use sparse embeddings.