Hugging Face TEI

¶Generate Embeddings

The TeiClient provides an embed() method to generate embeddings for input text. Embeddings are vector representations of text that capture semantic meaning, making them useful for tasks like semantic search, clustering, recommendation systems, and more.

Tip

Embeddings are numerical vectors that represent the meaning of text in a multi-dimensional space. Similar meanings will have closer vectors.

¶Example: Generate Embeddings

This example demonstrates how to generate a 1024-dimensional embedding for a given sentence.

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

// Load TEI endpoint for embeddings and API key
$teiUrl = getenv('TEI_EMBEDDINGS_URI');
$apiKey = getenv('TEI_API_KEY');

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

try {
    // Generate embeddings
    $embedding = $client->embed(inputs: "My name is Olivier and I");
    
    // Display the embedding
    var_dump($embedding);
} catch (MistralClientException $e) {
    echo $e->getMessage();
}

¶Output

array(1) {
  [0]=>
  array(1024) {
    [0]=>
    float(-0.012737735)
    [1]=>
    float(-0.0062732846)
    ...
    [1023]=>
    float(-0.017114446)
  }
}

• The result is an array of vectors, where:
  • Each vector represents the input text.
  • In this example, the vector is 1024-dimensional.

Note

The dimensionality (e.g., 1024) depends on the underlying embedding model served by the TEI endpoint.

¶Use Cases

• Semantic Search: Compare embeddings to find similar documents or queries.
• Clustering: Group similar texts together based on vector proximity.
• Recommendation Systems: Suggest similar items based on embedding similarity.
• Text Similarity: Measure the similarity between different texts using vector distance (e.g., cosine similarity).

¶Differences Between embed() and predict()

• Use embed() for semantic vector representations.
• Use predict() for structured insights like classification or sentiment analysis.

Warning

Always ensure the TEI endpoint URL is configured correctly for the type of task (embedding generation vs. prediction). Using the wrong endpoint may result in errors or unexpected outputs.