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.