Hugging Face TEI

¶Predict

The TeiClient not only supports generating embeddings but also allows you to run inference tasks such as sentiment analysis or emotion detection using the predict() method.

Tip

Use predict() when you want to classify text or extract structured predictions (e.g., sentiment, emotion, topic classification), depending on the underlying model served by the TEI endpoint.

¶Example: Sentiment/Emotion Analysis

This example demonstrates how to use the predict() method to analyze the emotions in a given sentence.

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

// Load TEI endpoint for sentiment analysis and API key
$teiUrl = getenv('TEI_SENTIMENT_ANALYSIS_URI');
$apiKey = getenv('TEI_API_KEY');

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

try {
    // Perform prediction
    $predictions = $client->predict(inputs: 'I love this product!');
    
    // Display results
    var_dump($predictions);
} catch (MistralClientException $e) {
    echo $e->getMessage();
}

¶Output

array(28) {
  [0]=>
  array(2) {
    ["score"]=>
    float(0.9895958)
    ["label"]=>
    string(4) "love"
  }
  [1]=>
  array(2) {
    ["score"]=>
    float(0.004673124)
    ["label"]=>
    string(10) "admiration"
  }
  ...
}

The output is an array of label-score pairs, where:

• label is the predicted category (e.g., emotion like "love", "joy").
• score is the confidence of the model in that label (between 0 and 1).

Note

The specific labels and their meanings depend on the model running at the TEI endpoint. In this example, the model predicts emotions.

¶Use Cases

• Emotion Detection: Classify emotions expressed in text (e.g., love, anger, joy).
• Sentiment Analysis: Determine whether a sentence expresses positive, negative, or neutral sentiment.
• Topic Classification: Depending on the model, classify text into different topics.

¶Error Handling

Always handle potential exceptions using try-catch to ensure robust applications.

try {
    $result = $client->predict(inputs: "sample input");
} catch (MistralClientException $e) {
    echo $e->getMessage();
}

Caution

Ensure that the TEI endpoint is configured to serve a model capable of prediction tasks (e.g., classification). Using the wrong endpoint may result in errors.

¶Differences Between embed() and predict()

Use embed() for semantic vector representations, and predict() for extracting structured insights like emotions or sentiments.