Support batch embeddings for Gemini provider

[larskoole]

Summary

The Gemini embeddings handler currently throws a PrismException when more than one input is provided via fromArray() or multiple fromInput() calls.

This adds support for multiple inputs by using the Gemini batchEmbedContents endpoint, which returns separate embedding vectors per input.

• Single input: continues using the existing embedContent endpoint (no behavior change)
• Multiple inputs: uses batchEmbedContents, mapping each embeddings[].values to an Embedding object

Changes

• src/Providers/Gemini/Handlers/Embeddings.php: Replaced the multi-input exception with routing logic. Added handleSingleRequest(), handleBatchRequest(), and sendBatchRequest() methods.
• tests/Providers/Gemini/EmbeddingsTest.php: Replaced the "throws exception" test with a batch embeddings test.
• tests/Fixtures/gemini/embeddings-batch-1.json: Fixture for the batch response.

Test plan

• Existing single-input embedding tests still pass
• New batch embeddings test verifies multiple inputs return separate embeddings
• All 7 Gemini embedding tests pass

[Copilot]

Pull request overview

Adds multi-input (batch) embedding support for the Gemini provider by routing requests with multiple inputs to Gemini’s batchEmbedContents endpoint, while keeping single-input behavior on embedContent.

Changes:

• Route Gemini embedding requests to embedContent (single input) vs batchEmbedContents (multiple inputs).
• Parse embeddings[].values from batch responses into Embedding objects.
• Update Gemini embeddings tests and add a batch response fixture.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
src/Providers/Gemini/Handlers/Embeddings.php	Adds single vs batch routing, batch request builder, and batch response mapping.
tests/Providers/Gemini/EmbeddingsTest.php	Replaces the “throws on multiple inputs” test with a batch embeddings test.
tests/Fixtures/gemini/embeddings-batch-1.json	Adds a recorded fixture representing a batchEmbedContents response.

Comments suppressed due to low confidence (1)

src/Providers/Gemini/Handlers/Embeddings.php:40

• Gemini error responses (e.g. {error: ...}) will currently be reported as a generic “Invalid response format...” because the handler never calls the shared ValidatesResponse logic used by other Gemini handlers. Consider applying Prism\Prism\Providers\Gemini\Concerns\ValidatesResponse and calling validateResponse($response) in both single and batch paths so provider error codes/messages aren’t lost.

        $response = $this->sendRequest($request);

        $data = $response->json();

        if (! isset($data['embedding'])) {
            throw PrismException::providerResponseError(
                'Gemini Error: Invalid response format or missing embedding data'
            );
        }

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

handleBatchRequest() only checks that embeddings exists, but will silently create empty Embedding objects when an item is missing values. This can hide partial/malformed provider responses; consider validating each item has a non-empty values array (and throwing a providerResponseError if not).

Suggested change

	if (! isset($data['embeddings'])) {
	throw PrismException::providerResponseError(
	'Gemini Error: Invalid batch response format or missing embeddings data'
	);
	}
	$embeddings = array_map(
	fn (array $embedding): Embedding => Embedding::fromArray($embedding['values'] ?? []),
	$data['embeddings'],
	);
	if (! isset($data['embeddings']) || ! is_array($data['embeddings'])) {
	throw PrismException::providerResponseError(
	'Gemini Error: Invalid batch response format or missing embeddings data'
	);
	}
	$embeddings = [];
	foreach ($data['embeddings'] as $index => $embedding) {
	if (! is_array($embedding)
	|| ! array_key_exists('values', $embedding)
	|| ! is_array($embedding['values'])
	|| $embedding['values'] === []
	) {
	throw PrismException::providerResponseError(
	sprintf(
	'Gemini Error: Invalid embedding item at index %d: missing or empty values',
	$index
	)
	);
	}
	$embeddings[] = Embedding::fromArray($embedding['values']);
	}

[Copilot]

In sendBatchRequest(), each inner request sets 'model' => "models/{$model}" while the single-request path uses $request->model() (no models/ prefix) and the endpoint URL already includes the model. This mismatch can cause the batch call to fail if the API expects the inner request model (when present) to exactly match the URL model; consider omitting the inner model field or using the same format as the single-request path.

[Copilot]

The new batch path isn’t covered beyond the happy-path mapping. Consider adding a test that exercises batch embeddings with provider options (e.g. title / taskType / outputDimensionality) to ensure those options are still sent/handled correctly when routing to batchEmbedContents.