Tablestore:Vector search

Scenarios

Vector search applies to recommendation systems, image and video retrieval, natural language processing, and semantic search.

• Retrieval-augmented generation (RAG)

  RAG is an AI framework that combines retrieval with Large Language Models (LLMs) to improve the accuracy of LLM outputs, especially for private or professional data. RAG is widely used in knowledge base scenarios.

• Recommendation system

  On e-commerce, social media, and streaming platforms, user behaviors, preferences, and content features are encoded into vectors. Vector search then finds matching products, articles, or videos, enabling personalized recommendations that improve user satisfaction and retention.

• Similarity detection (images, videos, and voice)

  In image, video, voice, voiceprint, and facial recognition, unstructured data is converted into vector representations. Vector search then finds the most similar targets. For example, when a user uploads an image on an e-commerce platform, the system finds product images with similar styles, colors, or patterns.

• Natural language processing and semantic search

  In natural language processing (NLP), text is converted into vector representations such as Word2Vec or BERT embeddings. Vector search then matches query semantics to find the most relevant documents, news, or Q&A pairs, improving search result relevance and user experience.

• Knowledge graph and AI chat

  Knowledge graph nodes and relationships can be represented as vectors. Vector search accelerates entity linking, relationship inference, and AI chat response times, enabling more accurate answers to complex questions.

Core advantages

Function overview

A KnnVectorQuery finds the most similar data items in a large-scale dataset by performing an approximate nearest neighbor query on numerical vectors.

Vector search inherits all the features of a search index. As a ready-to-use, pay-as-you-go service, it requires no system deployment. Stream-based index building allows data to be queried in near-real-time after it is written. It also supports high-throughput additions, updates, and deletions, with query performance comparable to HNSW-based systems.

To use KnnVectorQuery, specify the query vector, the vector field to search, and the number of nearest neighbors (TopK) to retrieve. The query returns the TopK vectors most similar to the query vector. You can combine it with other non-vector search features to filter results as needed.

Notes

Note the following when using vector search:

• Limits apply to the number of vector field types, dimensions, and other properties. For more information, see Search index limits.

• The search index is partitioned on the server side. Each partition returns its own TopK nearest neighbors, and the results are then aggregated at the client node. Therefore, if a token is used to paginate through all data, the total number of rows returned depends on the number of server-side partitions.

• Currently, the vector search feature is available in the following regions: China (Hangzhou), China (Shanghai), China (Qingdao), China (Beijing), China (Zhangjiakou), China (Ulanqab), China (Shenzhen), China (Guangzhou), China (Chengdu), China (Hong Kong), Japan (Tokyo), Singapore, Malaysia (Kuala Lumpur), Indonesia (Jakarta), Philippines (Manila), Thailand (Bangkok), Germany (Frankfurt), UK (London), US (Virginia), US (Silicon Valley), SAU (Riyadh - Partner Region).

Procedure

1. Generate vectors.

   Use open source models to convert data in Tablestore into vectors and store them.

2. Write vector data to Tablestore.

3. When you create a search index, configure the vector field.

   Configure the vector field's type, dimensions, and distance measure algorithm.

4. Use vector search to query data.

Billing

During the public preview, you are not charged for billable items specific to the KNN vector query feature. You are charged for other billable items based on existing billing rules.

When you use a search index to query data, you are charged for the read throughput consumed. For more information, see Billable items of search indexes.

Appendix 1: Use with BoolQuery

KnnVectorQuery and BoolQuery can be combined in different ways. The following two examples assume a scenario where a filter matches a small amount of data.

Assume a table contains 100 million images. A user has 50,000 images in total, but only 50 were added in the last 7 days. The user wants to find the 10 most similar images from those added in the last 7 days using search by image. The following table shows how these two combination methods differ.

Appendix 2: Vector normalization example

The following code normalizes a vector:

  public static float[] l2normalize(float[] v, boolean throwOnZero) {
    double squareSum = 0.0f;
    int dim = v.length;
    for (float x : v) {
      squareSum += x * x;
    }
    if (squareSum == 0) {
      if (throwOnZero) {
        throw new IllegalArgumentException("can't normalize a zero-length vector");
      } else {
        return v;
      }
    }
    double length = Math.sqrt(squareSum);
    for (int i = 0; i < dim; i++) {
      v[i] /= length;
    }
    return v;
  }