Autogenerating embeddings for Vector Search 2.0

You can use Vector Search 2.0 to autogenerate embeddings for your Collections. This lets you build new embeddings and deploy them instantly, which streamlines the path from raw data to a live production-scale search engine.

Supported Vertex AI embedding models

Note: Support for embedding models is subject to Vertex AI versioning and lifecycle guidelines. For more information, including retirement dates, see Model versions and lifecycle.

Vector Search 2.0 supports the following embedding models:

• Gemini — Provides state-of-the-art performance for embedding text (English-only and multilingual).

• Text — Specializes in (English-only and multilingual) and source code data.

 

Note: Support for multimodal embedding (multimodalembedding@001) is coming soon.

 

The following table provides details on each supported model.

Model	Description	Max output dimensions	Max sequence length (tokens)	Supported modalities and text languages	Additional limits
gemini-embedding-001	State-of-the-art performance across English, multilingual and code tasks. It unifies the previously specialized models like text-embedding-005 and text-multilingual-embedding-002 and achieves better performance in their respective domains.	3072	2048	Supported text languages	Embedding limits
gemini-embedding-2-preview	This is a next-generation multimodal embedding model from Google. Built on on the latest Gemini model architecture, this "omni embedding model" maps text, image, video, and PDF data into a single, unified embedding space.	3072	8192	Interleaved text, image, video, and PDF	API limits
text-embedding-004	Specialized in English and code tasks.	768	2048	English	API limits
text-embedding-005	Specialized in English and code tasks.	768	2048	English	API limits
text-multilingual-embedding-002	Specialized in multilingual tasks.	768	2048	Supported text languages	API limits

 

Note: For gemini-embedding-001 technical details, go to Gemini Embedding: Generalizable Embeddings from Gemini.

Creating Collections with autogenerated embeddings

When creating a Collection, specify the embedding model in the model_id field of vertex_embedding_config. This model is used whenever a Data Object is created without genre_embedding data defined.

The following code demonstrates how to specify the embedding model to use when autogenerating embeddings.

request = vectorsearch.CreateCollectionRequest(
   parent=f"projects/{PROJECT_ID}/locations/{LOCATION}",
   collection_id=collection_id,
   collection={
       "data_schema": {
           "type": "object",
           "properties": {
               "year": {"type": "number"},
               "genre": {"type": "string"},
               "director": {"type": "string"},
               "title": {"type": "string"},
           },
       },
       "vector_schema": {
           "plot_embedding": {"dense_vector": {"dimensions": 3} },
           "soundtrack_embedding": {"dense_vector": {"dimensions": 5} },
           "genre_embedding": {
               "dense_vector": {
                   "dimensions": 4,
                   "vertex_embedding_config": {
                       # If a data object is created without a supplied value for genre_embedding, it will be
                       # auto-generated based on this config.
                       "model_id": "text-embedding-004",
                       "text_template": ("Movie: {title} Genre: {genre} Year: {year}"),
                       "task_type": "RETRIEVAL_DOCUMENT",
                   },
               }
           },
           "sparse_embedding": {"sparse_vector": { } },
       },
   },
)
operation = vector_search_service_client.create_collection(request=request)
operation.result()

In the example code, a new Collection is created with the model_id field set to text-embedding-004. See Supported Vertex AI embedding models for which embedding models can be specified for model_id.

Quotas

Autogenerated embeddings rely on on customer quotas for the underlying Vertex AI embedding models. This is primarily constrained by two main quotas:

• Embed content input tokens per minute per region per base_model.

• Online prediction requests per base model per minute per region per base_model.

Make sure you have enough quota before creating Data Objects or running an import job.

See Manage your quota using the console for information on how to request larger quotas.

See also

• Collections
• Data Objects