SQL Database Vector Search

Vector Search is an Azion SQL Database feature that enables customers to implement semantic search engines. While traditional search models aim to find exact matches, such as keyword matches, vector search models use specialized algorithms to identify similar items based on their mathematical representations, or vector embeddings.

By using Vector Search, you can implement various use cases:

• Enhancing search systems and offering personalized recommendations by finding items with similar characteristics or based on users’ preferences, such as related products in ecommerce or content in streaming platforms.
• Creating text embeddings to search for semantically similar text, where words or phrases are represented as vectors.
• Building AI-based applications, leveraging Natural Language Processing (NLP) for voice assistants and chatbots, with integration to AI Inference that enables powerful Retrieval-Augmented Generation (RAG) applications using frameworks like LangChain and LangGraph directly at the edge.

Distributed across the Azion global edge network, this feature enables more relevant search results, real-time recommendations, and insights, drastically reducing latency and improving user satisfaction. All of this while maintaining data locality and reducing dependence on the centralized database.

Implementation

Databases and storage

By leveraging SQL Database, vector search databases are optimized to handle high-dimensional vector data at the edge. This enables fast, localized processing as well as reduced latency, allowing complex tasks for advanced data-intensive applications to run efficiently.

SQL Database implements Main/Replicas, distributed within the Azion Global Infrastructure, to enable ultra-low latency querying at the edge. This approach allows it to be accessed from any edge location, facilitating real-time processing and data analysis, and guaranteeing availability and fault tolerance. SQL Database uses SQLite’s dialect.

Columns

To store vectors in a vector search database, you can add a column specifically for the vector data.

For example, using the text-embedding-3-small model and a 1536 dimension, this column should be declared to hold an array of 32-bit floating-point numbers as a binary large object (BLOB) type, as follows. The (3) in the example specifies the number of 32-bit floating-point (F32) elements in the vector, indicating a 3-dimensional vector:

Then, you can insert data in the table, including vector embeddings (in this example, representing team stats for the 2023 season):

Embeddings

Embeddings are numerical vector representations of complex data (like words or images) that capture essential characteristics, enabling similarity-based searches. In the example, given an embedding of [80, 30, 60] for a team, a query can retrieve other teams with similar embeddings, helping identify teams with comparable performance stats.

Using embeddings, you can query to find similar information between the teams. For example, teams with stats similar to 82 goals scored, 25 goals conceded, and 63% possession:

Indexing

Considering Vector Search uses larger databases and datasets, it supports indexing through Approximate Nearest Neighbors (ANN), using SQL, wrapping the vector column into the libsql_vector_idx function.

To adequately use the index, you can modify your query to guarantee the index is consulted. Using the index is not automatic, since it is internally represented as a different table. For the previous example query, we can modify it slightly to make sure the index is consulted:

Limits and considerations

• Vector Search uses libSQL.
• LibSQL uses the native SQLite BLOB storage class for vector columns, defined according to the number of bits used to represent each floating-point number in the vector.
• Supported vector types: FLOAT1BIT | F1BIT_BLOB, FLOAT8 | F8_BLOB, FLOATB16 | FB16_BLOB, FLOAT16 | F16_BLOB, FLOAT32 | F32_BLOB, and FLOAT64 | F64_BLOB. The FLOAT32 is recommended as a starting point.
• The vector_distance_cos function calculates the cosine distance, which is defined at: Cosine Distance = 1 — Cosine Similarity.
  • The cosine distance ranges from 0 to 2, where:
    • A distance close to 0 indicates that the vectors are nearly identical or exactly matching.
    • A distance close to 1 indicates that the vectors are orthogonal (perpendicular).
    • A distance close to 2 indicates that the vectors are pointing in opposite directions.
• Euclidian distance isn’t supported for 1-bit FLOAT1BIT vectors.
• LibSQL can only operate on vectors with no more than 65536 dimensions.
• LibSQL implements the DiskANN algorithm to speed up approximate nearest neighbor queries for tables with vector columns. Check the Indexing section for more information.
• To query a vector index, you must use the vector_top_k(idx_name, q_vector, k) function, which finds the top k nearest neighbors and returns their ROWID or PRIMARY KEY.
• Vector index works only for tables with ROWID or with singular PRIMARY KEY. Composite PRIMARY KEY without ROWID isn’t supported.