Vector Search 2.0 interactive demo

This interactive demo lets you explore Vector Search 2.0, the fully managed vector database for Google Cloud. Use the demo to search across millions of product listings with semantic search (powered with Gemini auto-embeddings keyword text search or hybrid search that uses Reciprocal Rank Fusion (RRF) to combine both semantic search and keyword text search. Results can be reranked which uses the Gemini Enterprise Agent Platform Ranking API for reranking.

Vector Search 2.0 benefits

Vector Search required separate embedding generation, index deployment, and a Vertex AI Feature Store for metadata. Vector Search 2.0 provides unified storage for data and embeddings, automatic embedding generation, and built-in text search without requiring you to manage infrastructure.

Try it!

To run:

  1. In the Query field, describe the items you want to query for (for example, vintage 1970s pinball machine). Alternatively, click Generate Query to auto-generate a description.

  2. Click Submit or press Enter.

Experiment with the different options in the demo to get a head start understanding Vector Search 2.0 and the basics of vector search technology. For more information, go to User Interface.

User interface

This section describes settings in the UI you can use to control the results Vector Search 2.0 returns and how they are ranked.

Dataset

Use the Dataset drop-down to choose which dataset Vector Search 2.0 will run your query against. For information about each dataset, go to Datasets.

Query

To run a query:

  1. Click Dataset and choose which dataset to run your query against.

  2. In the Query field, add a description, ask a question, or enter one or more keywords to specify the items Vector Search 2.0 to return results for. Alternatively, click Generate Query to auto-generate a query.

  3. Do one of the following:

    1. To return semantically similar results, select Semantic.
    2. To return results based on the query's text syntax, select Text.
    3. To perform a hybrid search, selectic both Semantic and Text.

  4. If you are semantically querying either the Mercari 3M items (128-dim Gemini text embeddings) or Mercari 3M items (768-dim Gemini text embeddings) dataset, click Query Task Type and select either QUESTION_ANSWERING or RETRIEVAL_QUERY.

  5. Click Submit or press Enter to run the query.

Modify

There are many options you can use to modify the results Vector Search 2.0 returns from a query.

  • Click Rows and choose the maximum number of search results that you want Vector Search 2.0 to return.

  • Select Semantic to return semantically similar results.

  • Select Text to return results based on your query's text syntax.

  • Select both Semantic and Text to enable hybrid search.

  • Select Rerank to rerank search results based on their query relevance. For more information, see Improve search and RAG quality with ranking API.

  • Select Use kNN to use the k-nearest neighbors algorithm to retrieve search results. Because kNN provides a 100% recall rate, it is very useful when debugging or testing.

  • Click Query Task Type and select the which query task type to use. This is only available when using the Mercari 3M items (128-dim Gemini text embeddings) and Mercari 3M items (768-dim Gemini text embeddings) datasets. For more information about query task types, go to Choose an embeddings task type.

  • In the RRF Alpha field, enter a value between 0.0 and 1.0 to specify the RRF ranking effect. This is only available for hybrid search (both Semantic and Text are selected).
    Alpha values of:

    • 1.0 - Rank results purely by semantic (dense) similarity.
    • 0.0 - Rank results purely by keyword (sparse) relevance.
    • 0.5 - Ranks results with requal weight is given to semantic (dense) similarity and keyword (sparse) relevance.

    For more information, see What is Reciprocal Rank Fusion?

Metrics

After a query runs, you are provided with latency metrics that breakdown the time it took for different stages of the search to complete.

This includes:

  • The total latency from all sources.
  • The amount of time it took to generate embeddings and perform a vector search.
  • The amount of time reranking took if Rerank is selected.
  • The total amount of latency from the network and other sources.

Embeddings

Vector Search 2.0 supports the following embedding types:

Query process

In Vector Search 2.0, the query process depends on the kind of searches being done and the embedding models.

Embedding generation and search happen in a single API call. Vector Search 2.0 auto-generates the query embedding server-side using the model specified in the vertex_embedding_config for the Collection, performs the search, and returns results with full metadata without needing a separate Vertex AI Feature Store fetch.

Word2Vec and multimodal

The client generates the query embedding and sends it to Vector Search 2.0 for vector search. The results still include full metadata from the Collection and a Vertex AI Feature Store fetch isn't needed.

Text

Vector Search 2.0 performs built-in keyword matching directly. Embeddings are not generated.

Datasets

Dataset Description Embedding Model Embedding Task Type
Mercari 1M items
(768-dim Gemini text embeddings)		 Uses the Gemini Embedding 2 embedding model with separate text and image embeddings. This dataset supports text-to-text and text-to-image search using different vector fields. gemini-embedding-2-preview SEMANTIC_SIMILARITY
Mercari 3M items
(128-dim Gemini text embeddings)		 Has a reduced number of embedding dimensions that provides a smaller index size and storage cost. This dataset may have lower accuracy than Mercari 3M items (768-dim Gemini text embeddings). gemini-embedding-001 RETRIEVAL_DOCUMENT
Mercari 3M items
(768-dim Gemini text embeddings)		 Has full-dimension embeddings with auto-embeddings and high text semantic accuracy. This dataset uses RETRIEVAL_DOCUMENT for asymmetric query-document matching. gemini-embedding-001 RETRIEVAL_DOCUMENT
Mercari 3M items
(768-dim Gemini text embeddings, semantic similarity)		 Uses the SEMANTIC_SIMILARITY task type for similarity matching. This dataset is the best for use cases that need "find similar items" functionality where query and documents have same semantics. gemini-embedding-001 SEMANTIC_SIMILARITY
Mercari 3M items
(1408-dim Vertex AI multimodal embeddings)		 Has Vertex AI multimodal embeddings that encodes item images for image search. multimodal-embedding-001 N/A
Mercari 3M items
(100-dim Word2Vec text embeddings)		 Uses the Gensim Word2Vec embedding model. This dataset has the fastest inference. It provides only word-level semantic understanding but is very compact. Gensim Word2Vec N/A

See also