Run a hybrid vector similarity search

Perform hybrid searches in AlloyDB for PostgreSQL, which combines text search, keyword matching, vector search, and semantic similarity, using the vector extension, which is a standard pgvector PostgreSQL extension customized for AlloyDB.

To improve search relevance, you can perform a hybrid search that combines vector similarity search with text search. AlloyDB lets you create Generalized Inverted Indexes (GIN) for text search and Scalable Nearest Neighbors (ScaNN) indexes for vector search. You can then combine results from both searches and re-rank them using reciprocal rank fusion (RRF), an algorithm that combines multiple search result lists into a single ranked list, to retrieve highly relevant results based on both exact keyword matches and semantic similarity.

To perform a hybrid search in AlloyDB for PostgreSQL, you'll need to create both a vector index and a text search index on your table. Then, you'll combine the results from both searches and re-rank them to present the most relevant information.

Create a GIN index

A Generalized Inverted Index (GIN) index is a specialized index type optimized for searching within composite values, such as arrays, JSONB, and full-text search data.

To create a GIN index on your text data to perform a full text search, run the following:

CREATE INDEX INDEX_NAME ON TABLE USING GIN (to_tsvector('english', COLUMN_NAME));

Replace the following:

  • INDEX_NAME: the name of the index you want to create —for example, my-gin-index.

  • TABLE: the table to add the index to.

  • COLUMN_NAME: the column that stores the text data you want to search.

Create a ScaNN index

To apply a two-level tree index using the ScaNN algorithm to a column containing stored vector embeddings, run the following DDL query:

CREATE INDEX INDEX_NAME ON TABLE
  USING scann (EMBEDDING_COLUMN DISTANCE_FUNCTION)
  WITH (num_leaves=NUM_LEAVES_VALUE);

Replace the following:

  • INDEX_NAME: the name of the index you want to create—for example, my-scann-index. The index names are shared across your database. Ensure that each index name is unique to each table in your database.

  • TABLE: the table to add the index to.

  • EMBEDDING_COLUMN: a column that stores vector data.

  • DISTANCE_FUNCTION: the distance function to use with this index. Choose one of the following:

    • L2 distance: l2

    • Dot product: dot_product

    • Cosine distance: cosine

  • NUM_LEAVES_VALUE: the number of partitions to apply to this index. Set to any value between 1 to 1048576. For more information about how to decide this value, see Tune a ScaNN index.

To learn more about different ScaNN index configurations, see Create a ScaNN index. You can also create a HNSW index.

Perform a hybrid search using reciprocal rank fusion

Hybrid search involves performing separate vector and text searches, then combining and re-ranking results using Reciprocal Rank Fusion (RRF). RRF is a rank-based algorithm that combines multiple ranked lists of search results into a single ranked list by assigning a score to each document. This score is based on RRF's reciprocal rank across all contributing lists, with higher-ranked documents receiving a greater contribution. Use the following SQL query to combine full text search and hybrid search, and re-rank the results:

        WITH vector_search AS (
            SELECT id,
                RANK () OVER (ORDER BY embedding <=> google_ml.embedding('MODEL_ID', 'TEXT')) AS rank
                FROM TABLE
                ORDER BY embedding <=> google_ml.embedding('MODEL_ID', 'TEXT') LIMIT 10
        ),
        text_search AS (
            SELECT id,
                RANK () OVER (ORDER BY ts_rank(to_tsvector('english', COLUMN_NAME), to_tsquery(KEYWORD)) desc)
            FROM TABLE
            WHERE to_tsvector('english', COLUMN_NAME) @@ to_tsquery(KEYWORD)
            ORDER BY ts_rank(to_tsvector('english', COLUMN_NAME), to_tsquery(KEYWORD)) desc
            LIMIT 10
        )
        SELECT
            COALESCE(vector_search.id, text_search.id) AS id,
            COALESCE(1.0 / (60 + vector_search.rank), 0.0) + COALESCE(1.0 / (60 + text_search.rank), 0.0) AS rrf_score
        FROM vector_search FULL OUTER JOIN text_search ON vector_search.id = text_search.id
        ORDER BY rrf_score DESC
        LIMIT 5;

Replace the following:

  • MODEL_ID: the ID of the model to query.

    If you are using the Vertex AI Model Garden, then specify text-embedding-005 as the model ID. These are the cloud-based models that AlloyDB can use for text embeddings. For more information, see Text embeddings.

  • TABLE: the table containing your data.

  • TEXT: the text to translate into a vector embedding.

  • KEYWORD: the keyword you want to search for.

  • COLUMN_NAME: a column that stores contains the text data you want to search.

Explanation of the Hybrid Search Query and related Common Table Expression (CTE):

  • vector_search CTE: Performs a standard vector similarity search, ordering results by cosine distance and assigning a rank. It retrieves the top 10 most semantically similar products.
  • text_search CTE: Executes a text search using to_tsvector and to_tsquery, calculating relevance with ts_rank and retrieving the top 10 most relevant text matches.
  • Final SELECT Statement CTE: Joins vector and text search results using a FULL OUTER JOIN, selects the product ID, calculates the RRF score, orders by score, and retrieves the top 5 results.

Perform a hybrid search using LangChain

Hybrid search with the AlloyDB vector store enhances search accuracy by combining two different lookup strategies: dense embedding vector search and keyword-based search. AlloyDBVectorStore is a LangChain vector store class that uses LangChain by acting as a specific implementation of LangChain's VectorStore class. Learn how to use AlloyDB to store vector embeddings with the AlloyDBVectorStore class.

You can enable and configure this hybrid search using the HybridSearchConfig class when you set up your AlloyDBVectorStore.

Hybrid search with the AlloyDB vector store simultaneously performs a semantic search to understand the meaning and context of a query, and a keyword search to find exact lexical matches. The results from both searches are then merged to provide a more comprehensive set of results.

What's next