VertexAI vectorizer uses deprecated model garden SDK (removal June 2026)

[automaton82]

VertexAIVectorizer._client calls TextEmbeddingModel.from_pretrained() and MultiModalEmbeddingModel.from_pretrained() from vertexai.language_models / vertexai.vision_models. These are deprecated as of June 24, 2025 and will be removed June 24, 2026.

Triggers on every vectorizer init:

UserWarning: This feature is deprecated as of June 24, 2025 and will be removed on June 24, 2026.
vertexai/_model_garden/_model_garden_models.py:278

Needs migration to the new google-genai SDK. Reproduced on 0.18.2.

[omribz156]

I traced the RedisVL side before trying a patch, and I think this is worth splitting rather than doing a broad blind migration.

Source trace:

• redisvl/utils/vectorize/vertexai.py currently initializes the Vertex AI SDK with vertexai.init(...).
• _client then imports TextEmbeddingModel from vertexai.language_models for text models and MultiModalEmbeddingModel from vertexai.vision_models for multimodal models.
• The text path uses get_embeddings([...]), while the multimodal path also depends on vertexai.vision_models.Image / Video for embed_image, embed_video, and cache serialization.

Google's migration guide says the deprecated Vertex AI SDK namespaces are removed after June 24, 2026 and shows the text embedding replacement through the Google Gen AI SDK models.embed_content API:
https://docs.cloud.google.com/vertex-ai/generative-ai/docs/deprecations/genai-vertexai-sdk#embeddings

The safer shape for RedisVL may be:

1. First migrate the text embedding path to google-genai, add google-genai to the vertexai extra, and cover it with mocked-client unit tests so no live GCP credentials are needed.
2. Treat multimodal separately. The current code exposes Image/Video objects and cache serialization around the old vertexai.vision_models types, while the current multimodal embeddings reference documents request/response fields such as imageEmbedding, textEmbedding, and videoEmbeddings:
   https://docs.cloud.google.com/vertex-ai/generative-ai/docs/model-reference/multimodal-embeddings-api

I can take a focused text-path PR if maintainers agree with splitting the migration that way. I did not open a PR yet because changing text + image + video behavior together would be easy to overreach without live-provider validation.

This trace was Codex-assisted, with the source and Google docs checked before posting.

[fkdosilovic]

I believe that we could do a more future-proof refactor that involves deprecating the google-cloud-aiplatform as a dependency and rely solely on google-genai as the main dependency for talking to Google APIs. The main benefits are:

1. The google-genai encapsulates both the Gemini Enterprise Agent Platform (formerly known as Vertex AI) and the Gemini API, and
2. The google-cloud-aiplatform has a serious issue with using additional 200 MB of memory (see This module is taking up too much memory. googleapis/python-aiplatform#6191). I stumbled upon this myself by accident and decided to avoid the dependency all together for now.

The refactor would involve:

1. adding a deprecation warning message to users that the VertexAIVectorizer would be deprecated in the near future (say version 0.21) and completely removed in version, say, 0.23. (same goes for VertexAITextVectorizer, which already is deprecated)
2. introduce a new optional group/dependnecy google-genai, separate of vertexai, and adding a new class GoogleGenAIVectorizer that implements the needed functionality.

The only thing I'm not sure about is whether to separate the multimodal capabilities into separate vectorizers or not, as @omribz156 suggested. I did not work with the images/video/audio that extensively, so will have to take a look at the google-genai interface more closely for that. From a quick look at the redis-vl-python source, I see that there is no general guideline in the API around whether the vectorizer should be multimodal or not.

I can also run integration tests for both the Gemini Enterprise Agent Platform and Gemini API. To make the tests more developer friendly, if the developer does not have exposed GEMINI_API_KEY or credentials for Gemini Enterprise Agent Platform, the corresponding integration tests can be skipped.

[vishal-bala]

Hey! Sorry for getting to this late, and thanks for flagging the issue. I'm a fan of your proposed approach @fkdosilovic - let's add a deprecation warning to the old vectorizer and introduce a new one for the google-genai dependency.

Given that the multimodal model is still public preview there, I'd be fine to split off just the text vectorizer for now and then migrate the multimodal functionality once it's either stable or requested.