[PredictionServiceClient] Bug in embedContent and EmbedContentRequest: Invalid model/endpoint mappings causing INVALID_ARGUMENT or undefined method errors

[Jony-Shark]

When attempting to use PredictionServiceClient::embedContent() with the newer EmbedContentRequest class, the underlying gRPC bindings fail to resolve the model or endpoint correctly. This makes it impossible to use the native text-embedding endpoints via the official PHP SDK.

Steps to reproduce:
If we try to construct the request by setting the model using setModel():

use Google\Cloud\AIPlatform\V1\Client\PredictionServiceClient;
use Google\Cloud\AIPlatform\V1\EmbedContentRequest;
use Google\Cloud\AIPlatform\V1\Content;
use Google\Cloud\AIPlatform\V1\Part;
 
$client = new PredictionServiceClient([
    'apiEndpoint' => 'us-central1-aiplatform.googleapis.com',
    'credentials' => '/path/to/credentials.json'
]);
 
$content = (new Content())->setParts([(new Part())->setText('Hello world')]);
 
$req = new EmbedContentRequest();
$req->setModel('projects/MY_PROJECT/locations/us-central1/publishers/google/models/text-embedding-004');
$req->setContent($content);
 
$res = $client->embedContent($req);

Actual Result: An ApiException is thrown from the backend:

{
    "message": "Invalid value (oneof), oneof field '_model' is already set. Cannot set 'model'",
    "code": 3,
    "status": "INVALID_ARGUMENT",
    "details": [ ... ]
}

If we try to supply an endpoint via the $optionalArgs to bypass the model binding:

$res = $client->embedContent($req, [
    'endpoint' => 'projects/MY_PROJECT/locations/us-central1/publishers/google/models/text-embedding-004'
]);

Actual Result:

Google\ApiCore\ValidationException: Could not map bindings for google.cloud.aiplatform.v1.PredictionService/EmbedContent to any Uri template.

If we attempt to use setEndpoint() on the request object itself (which is available on PredictRequest but seems to be missing here):

$req->setEndpoint('projects/MY_PROJECT/locations/us-central1/publishers/google/models/text-embedding-004');

Actual Result:

Fatal error: Call to undefined method Google\Cloud\AIPlatform\V1\EmbedContentRequest::setEndpoint()

Expected Result:
The EmbedContentRequest should properly bind the model parameter to the gRPC URI template (e.g. /v1/{model=projects//locations//publishers//models/}:embedContent) without throwing a protobuf oneof conflict (_model vs model), and should successfully return the EmbedContentResponse.

Workaround (for others facing this issue):

Currently, the only way to get embeddings via PHP is to bypass EmbedContentRequest and use the older PredictRequest with raw Protobuf Structs:

$instanceValue = new \Google\Protobuf\Value();
$struct = new \Google\Protobuf\Struct();
$struct->setFields(['content' => (new \Google\Protobuf\Value())->setStringValue('Hello World')]);
$instanceValue->setStructValue($struct);
 
$request = (new PredictRequest())
    ->setEndpoint('projects/MY_PROJECT/locations/us-central1/publishers/google/models/text-embedding-004')
    ->setInstances([$instanceValue]);
 
$response = $client->predict($request);

Environment details:
OS: Linux
PHP version: 8.0+
Package name and version: google/cloud-ai-platform (Tested on versions ^1.13.0 up to 1.60.1)

[szepeviktor]

Root Cause

This looks like an upstream proto definition bug, not a PHP-only implementation bug.

Upstream reference:
https://github.com/googleapis/googleapis/blob/master/google/cloud/aiplatform/v1/prediction_service.proto#L894

In google/cloud/aiplatform/v1/prediction_service.proto, EmbedContentRequest.model appears to be wrong in two ways:

• it is generated as optional string model = 1
• its resource_reference points to aiplatform.googleapis.com/Endpoint

That does not match the actual EmbedContent HTTP binding, which expects a publisher model:

post: "/v1/{model=projects/*/locations/*/publishers/*/models/*}:embedContent"

Because model is optional, the generated protobuf descriptor gets proto3 presence and a synthetic oneof (_model). In PHP this seems to collide with the REST URI binding for the same field, which explains errors like:

Invalid value (oneof), oneof field '_model' is already set. Cannot set 'model'

So the real issue is likely:

• model should not be proto3-optional here
• model should reference the publisher-model resource type, not Endpoint
• content also appears to be modeled too loosely, even though it is semantically required

The endpoint workaround is not valid for this method, because EmbedContent is bound on model, not endpoint.