Xai

¶Unified Embeddings API

This document explains the unified embeddings API exposed by the application. It describes:

• How to request embeddings in a provider-agnostic way (same methods across supported clients).
• The EmbeddingCollection and Embedding classes, including their public methods and expected behavior.
• Error handling and batching behavior.

The goal is to keep a consistent developer experience across providers while returning a normalized in-memory representation of embeddings.

Here is a list of compatible clients:

• Mistral AI La plateforme
• HuggingFace TEI (Text Embeddings Inference)
• LlamaCPP
• Ollama
• Vllm
• Voyage
• Xai

¶Supported scope

Any client that implements the unified API exposes the same methods to:

• Embed a single text.
• Embed multiple texts (optionally in batches).
• Produce a normalized EmbeddingCollection with Embedding objects.

Notes:

• The concrete client is responsible for calling the provider’s embeddings endpoint and for filling vectors into the returned EmbeddingCollection.
• The unified API hides provider request/response specifics from your application code.

¶Quick start

¶Single text

<?php
// Instantiate the client of your choice that supports the unified embeddings API.
// Replace placeholders with your values.
$client = /* new ...Client( ... ) */;

try {
    $collection = $client->embedText("What is the best French cheese?");
    foreach ($collection->getAll() as $embedding) {
        $id     = $embedding->getId()->toString();
        $text   = $embedding->getText();   // may be null depending on the provider flow
        $vector = $embedding->getVector(); // array of floats
        // use the vector as needed
    }
} catch (Throwable $e) {
    // handle error
}

¶Multiple texts (optional batching)

<?php
$client = /* new ... */;

$texts = [
    "What is the best French cheese?",
    "How to make Dijon mayonnaise?",
];

// batch: pass null to defer to client defaults, or an integer to force chunk size.
$collection = $client->embedTexts($texts, model: null, batch: 32);

// You can iterate directly (EmbeddingCollection is iterable)
foreach ($collection as $embedding) {
    // ...
}

¶Embeddings from with an EmbeddingCollection object

<?php
use Partitech\PhpMistral\Clients\Mistral\MistralClient;
use Partitech\PhpMistral\Embeddings\EmbeddingCollection;
use Partitech\PhpMistral\Exceptions\MistralClientException;

// export MISTRAL_API_KEY=your_api_key
$apiKey = getenv('MISTRAL_API_KEY');
$client = new MistralClient($apiKey);

$inputs = [];

for($i=0; $i<10; $i++) {
    $inputs[] = "$i : What is the best French cheese?";
}

$embeddingCollection = (new EmbeddingCollection())
    ->setModel('mistral-embed')
    ->fromList($inputs)
    ->setBatchSize(3);
$result = $client->createEmbeddings($embeddingCollection);

¶Public API (unified methods)

These methods are made available on compatible clients through a shared trait.

¶embedText(string $text, ?string $model = null): EmbeddingCollection

• Purpose: embed a single text.
• Behavior: returns an EmbeddingCollection containing one Embedding. The Embedding will have a generated UUID and its vector filled by the concrete client.
• Model: optionally override the model to use (client-specific).
• Errors: throws an exception if the input text is empty.

¶embedTexts(array $texts, ?string $model = null, int $batch = null): EmbeddingCollection

• Purpose: embed multiple texts at once.
• Behavior:
  • Builds an EmbeddingCollection from the given texts.
  • If $batch is provided, it sets the batch size to guide chunking on the client side.
  • Returns a new EmbeddingCollection with vectors filled.
• Model: optionally override the model to use (client-specific).
• Errors: throws an exception if the input list is empty or its first element is empty.

¶setBatchSize(?int $batchSize): self and getBatchSize(): ?int

• Purpose: configure/read the default batch size on the client instance exposing the trait.
• Note: batching is effectively applied by the concrete client when building requests.

¶EmbeddingCollection class

A container for Embedding objects. It is iterable and keyed by Embedding UUIDs (strings).

¶Properties (conceptual)

• model: string
• batchSize: ?int
• embeddings: array<string, Embedding> (internal map keyed by UUID string)

¶Construction/Population helpers

• add(Embedding $embedding): void
  • Adds an Embedding to the collection (keyed by its UUID).
• setAll(array $embeddings): void
  • Replaces the collection with the given array of Embedding.
  • Throws InvalidArgumentException if any entry is not an Embedding.

¶Lookup and access

• getAll(): array
  • Returns the internal array of Embedding keyed by UUID.
• getById(string $id): ?Embedding
  • Returns the Embedding for a UUID, or null.
• removeById(string $id): void
  • Removes the Embedding by UUID.
• getByIds(array $ids): ?EmbeddingCollection
  • Returns a new EmbeddingCollection with items matching the provided UUID list.
• count(): int
  • Number of embeddings in the collection.
• getIterator(): Traversable
  • Enables iteration over the collection (Embedding objects).

¶Positional helpers

• getIdByPos(int $pos): string
  • Returns the UUID at the given zero-based position.
• getByPos(int $pos): Embedding
  • Returns the Embedding at the given position (via UUID).
• updateVectorsByPos(int $pos, array $vectors): void
  • Replaces the vector of the Embedding located at the given position.

¶Batching helpers

• chunk(?int $batchSize = null): array<EmbeddingCollection>
  • If $batchSize is null, returns an array with this collection only.
  • Otherwise, splits into an array of EmbeddingCollection chunks, preserving model and batchSize.

¶Normalization helpers

• arrayAsEmbeddingCollection(array $result): ?EmbeddingCollection
  • Builds and returns a new EmbeddingCollection by reading entries shaped like:
    • ['text' => ?string, 'vector' => array, 'id' => ?string]
  • Preserves current model and batchSize on the new collection.
• fromList(array $result): ?EmbeddingCollection
  • Convenience: converts a string list into a collection of items like ['text' => $value], then delegates to arrayAsEmbeddingCollection.

¶Model and batching metadata

• getModel(): string
• setModel(?string $model): EmbeddingCollection
  • Sets the model name to be associated with this collection. Returns self for chaining.
• getBatchSize(): ?int
• setBatchSize(?int $batchSize): EmbeddingCollection
  • Sets an optional batching hint. Returns self for chaining.

¶Embedding class

Represents a single text embedding payload.

¶Properties (conceptual)

• id: UuidInterface
• text: ?string
• vector: array

¶Constructor

• __construct(?string $text = null, array|string $vector = [], null|string|UuidInterface $id = null)
  • If $id is null or invalid, a new UUID v4 is generated.
  • $vector may be provided as:
    • array of numbers, or
    • a string formatted like "[...]" or a CSV-like list; it will be parsed into an array.

¶Accessors/Mutators

• getId(): UuidInterface
• getText(): ?string
• setText(?string $text): self
• getVector(): array
• setVector(string|array $vector): self
  • Accepts an array or a string; if string, trims surrounding brackets and splits by comma.
• setId(null|string|UuidInterface $id): self
  • If a string is passed, it is parsed as a UUID.
  • If null or not a UUID, a new UUID v4 is generated.

¶Error handling

• Single text: calling embedText with an empty string throws an exception.
• Multiple texts: calling embedTexts with an empty array or with an empty first item throws an exception.
• Provider-side failures (network, model, quota, input format) are surfaced by the concrete client when executing the request during createEmbeddings.

¶Batching behavior

• If you pass a $batch value to embedTexts, it is copied to the EmbeddingCollection and used by the concrete client to split requests.
• chunk(...) on EmbeddingCollection returns arrays of EmbeddingCollection chunks, each preserving model and batchSize. This can be leveraged by clients or by advanced consumers.

¶Output shape (normalized)

• The returned EmbeddingCollection aggregates Embedding instances keyed by UUID.
• Each Embedding contains:
  • id: UUID (string via toString()).
  • text: the original text when available.
  • vector: the embedding vector as an array.

This normalized shape allows consistent handling of embeddings regardless of the provider used by the concrete client.