Multimodal embedding via genai SDK fails

[Attila-Roomr]

Using Python:
We are migrating to the google-genai SDK (v1.x) and targeting the multimodalembedding model on Vertex AI. The SDK's models.embed_content method incorrectly wraps inputs in the Gemini contents/parts schema, which the legacy multimodalembedding backend rejects with 400 Empty instances. This persists regardless of whether we use types.Part, types.Content, or raw dictionaries as inputs.
I've tried all the solutions that Gemini was able to suggest, they all failed for various reasons, but most notably the Empty instance was the most recurring one.

The vertex ai SDK works perfectly, but we were asked to migrate our 1408 multimodal embedding code.

[Attila-Roomr]

I was also asked by Gemini to pass this on:

Issue: Incompatibility between google-genai SDK (v1.x) and Vertex AI multimodalembedding model.

Root Cause: The SDK's models.embed_content method serializes inputs using the Gemini-style {"contents": [{"parts": [...]}]} schema. However, the multimodalembedding backend is a legacy Vertex AI model that strictly requires a flat {"instances": [...]} array.

Symptoms: > 1. Passing types.Part or types.Content objects results in 400 Empty instances because the backend finds no data in the instances root. 2. Passing raw dictionaries to bypass nesting results in Pydantic validation errors (_EmbedContentParameters) as the SDK forces specific types that do not match the legacy schema. 3. Standard workarounds like client.models.predict or client.api are either missing or restricted in the current GA version of the SDK.

Technical Summary for Google Support
Product: Vertex AI / google-genai Python SDK (v1.x) Model: multimodalembedding

Issue Description: > The google-genai SDK is unable to successfully call the multimodalembedding model on Vertex AI, consistently returning a 400 Empty instances error.

Root Cause Analysis:

Serialization Mismatch: The SDK's models.embed_content method attempts to format the request using the Gemini {"contents": [{"parts": [...]}]} schema. However, the multimodalembedding model is a legacy Vertex AI model that strictly requires a flat {"instances": [...]} array.

Backend Rejection: Because the data is nested inside a parts key, the legacy backend fails to find the input data in the root of the "instances" object, triggering the "Empty instances" response.

Validation Lock: Attempting to pass raw dictionaries to correct the structure manually results in Validation errors for _EmbedContentParameters because the SDK's Pydantic models enforce the Gemini-style schema.

Missing Accessors: Known "escape hatch" attributes like client.api_client, client.vertex, or client.request appear to be private, restricted, or renamed in the current GA version, preventing a manual override within the SDK's native transport.

Reproduction Code:

Python
response = client.models.embed_content(
model='multimodalembedding',
contents=[types.Part.from_text("test"), types.Part.from_bytes(data=b"...", mime_type="image/jpeg")]
)
'# Result: google.api_core.exceptions.InvalidArgument: 400 Empty instances.

[janasangeetha]

Hey @Attila-Roomr
Thanks for reaching out!
As Generative AI module in the Vertex AI SDK is deprecated and will no longer be available after June 24, 2026. You need to replace the Vertex AI SDK implementation with the Google Gen AI SDK. Please find the example for multimodal embeddings.
Please share runnable code incase you are still facing issues.

[steve-mavens]

@janasangeetha : The URL you provided only talks about text embeddings, not multimodal embeddings. Do you have any runnable code to demonstrate using the genai SDK to get multimodal embeddings? Specifically the multimodalembedding@001 model?

Google's docs provide example code using the vertex_ai SDK here: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/embeddings/get-multimodal-embeddings#text-image-embedding , so I think the question is how to migrate that to genai.