Tablestore:How do I optimize the performance of Tablestore KNN vector query?

1. Check the sorting method

Check whether the sorting method is set to ScoreSort to sort the query results by score when you use the KNN vector query feature. By default, the query results are sorted by primary key.

2. Adjust the combination of KnnVectorQuery and BoolQuery

When you use the combination of KnnVectorQuery and BoolQuery, we recommend that you set the query type of the search index to KnnVectorQuery. If you include the query conditions of BoolQuery in the KNN vector query filter (Filter), the scoring is not affected.

The following sample Java code provides an example on how to use the KNN vector query feature:

private static void knnVectorQuery(SyncClient client) {
    SearchQuery searchQuery = new SearchQuery();
    KnnVectorQuery query = new KnnVectorQuery();
    query.setFieldName("Col_Vector");
    query.setTopK(10); // Return the top 10 vectors in the table that have the highest similarity to the specified vector. 
    query.setFloat32QueryVector(new float[]{0.1f, 0.2f, 0.3f, 0.4f});
    // Specify the query conditions for the top 10 vectors: the value of the Col_Keyword column is hangzhou and the value of the Col_Long column is less than 4. 
    query.setFilter(QueryBuilders.bool()
            .must(QueryBuilders.term("Col_Keyword", "hangzhou"))
            .must(QueryBuilders.range("Col_Long").lessThan(4))
    );
    searchQuery.setQuery(query);
    searchQuery.setLimit(10);
    // Sort the query results by score. 
    searchQuery.setSort(new Sort(Collections.singletonList(new ScoreSort())));
    SearchRequest searchRequest = new SearchRequest("<TABLE_NAME>", "<SEARCH_INDEX_NAME>", searchQuery);
    SearchRequest.ColumnsToGet columnsToGet = new SearchRequest.ColumnsToGet();
    columnsToGet.setColumns(Arrays.asList("Col_Keyword", "Col_Long"));
    searchRequest.setColumnsToGet(columnsToGet);
    // Call the Search operation. 
    SearchResponse resp = client.search(searchRequest);
    for (SearchHit hit : resp.getSearchHits()) {
        // Display the scores. 
        System.out.println(hit.getScore());
        // Display the data. 
        System.out.println(hit.getRow());
    }
}

3. Check the vector generation performance

Tablestore calculates the similarity between vectors and does not consider whether the vector generation performance is optimal. The vectors in databases and the vectors that you want to query are generated by external Embedding models. In specialized fields, the generated vectors may be inaccurate. To troubleshoot poor vector generation performance, perform the following steps:

1. Use methods or tools other than Tablestore for scoring.

   1. Name the vector that you want to query as Vector a and the vector that you want to return from a Tablestore table as Vector b.

      Note

      You can use a search index or a secondary index or the operation for reading data from a data table in the Wide Column model to obtain the data of Vector b.

   2. Calculate Score a, which represents the similarity between Vector a and Vector b, by using the MetricFunction.COSINE.compare(a, b) method described in Appendix: Sample code for the scoring formulas of KNN vector query.

2. Use Tablestore for scoring.

   Use the KNN vector query feature of Tablestore to query Vector a and check Score b of each row in the response.

3. Compare and analyze the scores.

   If no rows that contain Vector b are found in the KNN vector query of Tablestore, Score b of each row in the response is theoretically higher than Score a.

   The scores indicate that the poor KNN vector query performance is caused by the poor vector generation performance of the Embedding model. The KNN vector query response includes vector data whose score is higher than the expected score. The vector data whose score is less than the expected score is not returned.

4. Select a solution.

   In most cases, the preceding issue occurs in specialized fields. For example, vector generation does not work as expected for terms in biotechnology and medical industries when general Embedding models are used. Specific terms may have similar semantics in specialized fields but significantly different semantics in general Embedding models. In this case, consider using the following solutions:

   • Use Embedding models dedicated for specialized fields.

     ModelScope provides various Embedding models. You can select an Embedding model for a specialized field, such as public service, e-commerce, healthcare, justice, or finance, based on your business requirements. For more information, see Embedding model list.

   • Collect a large amount of specialized corpus by using legal methods to train a suitable Embedding model.