Tuning Vector Indexes and Queries

When to Use a Vector Index

SingleStore recommends using vector indexes for searches over very large sets of vectors (e.g. billions of vectors) when there are no other filters in the query that can be used to reduce the number of vectors being searched. That is, if your data set is only millions of vectors, or if there is a WHERE clause (filter) in the query which will eliminate many of the rows in the table, you may have acceptable performance without using an index. As vector indexes use memory, if you can achieve acceptable performance without an index, SingleStore recommends not creating an index and using brute-force (scan) for your queries.

Types of Vector Indexes

SingleStore supports several types of vector indexes, but recommends using IVF_PQFS and HNSW_FLAT. Refer to Vector Indexing for more information on types of indexes and configuration parameters supported by SingleStore.

Example Table

The following example table represents product reviews from a web site. The table contains the text of the reviews as well as a vector embedding that captures the meaning of the review. Embeddings are a common use of vectors as is described in Working with Vector Data. The columns of the example table are:

• id: an INT id

• review: a TEXT field which stores the text of the comment

• review_embedding: a VECTOR of 1024 dimensions which stores a vector embedding of the review

• category: a VARCHAR which stores a category assigned to the review

Below is the SQL that creates the example reviews table.

CREATE TABLE reviews (
  id INT OPTION 'Integer',
  review TEXT OPTION 'SeekableString',
  review_embedding VECTOR(1024) NOT NULL OPTION 'SeekableString',
  category VARCHAR(256));

Note that the id column is qualified OPTION 'Integer', and the review and review_embedding columns are qualified as OPTION 'Seekable String'. The OPTION clauses specify how SingleStore should encode the data in those columns and will improve query performance.

The reviews table will be used to demonstrate creating and tuning indexes and to discuss advantages, disadvantages, and tradeoffs during index creation and tuning.

General Advantages and Disadvantages

There are several index types and a variety of configuration parameters. Vector indexes can be chosen and tuned to provide the accuracy, memory size, and search time that your application requires.

In general, IVF_PQFS has lower index build time and index size and HNSW_FLAT has lower search time and higher accuracy (recall). IVF_PQFS uses a fraction of the memory (~10x reduction or more) of HNSW and has a faster index build. The disadvantage of IVF_PQFS is potentially lower accuracy (recall) and longer search time; however, the recall and search time of IVF_PQFS can be improved by tuning the index building and search parameters as described in the section Tune IVF_PQFS Parameters below.

HNSW gives excellent performance (search time) and accuracy (recall); however, the size of an HNSW_FLAT index is the total size of the vectors. SingleStore keeps vector indexes in memory, so to use HNSW_FLAT, you will need to have enough memory to store the index.

As the size of an IVF_PQFS index is significantly smaller than an HNSW_FLAT index, tuning an IVF_PQFS index to give good performance and recall can be a good solution for many situations. SingleStore recommends IVF_PQFS unless your application requires high recall and very fast search speed; use HNSW_FLAT only if you cannot obtain good enough performance (search time) and accuracy (recall) by tuning IVF_PQFS.

Create the Vector Index

SingleStore recommends starting with an IVF_PQFS index due to the reduced memory required to store IVF_PQFS indexes. When building the IVF_PQFS index, SingleStore recommends setting the index building parameter m to be the number of dimensions of the vector column divided by 4.

The following command creates such an index for the reviews table, where m is set to 256 (which is 1024/4) and nlist is set to 1024 and nprobe is set to 20. Refer to nlist and nprobe for more information on setting nlist and nprobe. The metric_type is set to DOT_PRODUCT, which is also the default.

The parameter m is an index building parameter that determines the number of subquantizers used in product quantization, a method for compressing vectors. The parameter nlist specifies the number of clusters created during index build. The parameter nprobe specifies the number of probes to be used at query time.

ALTER TABLE reviews
ADD VECTOR INDEX ivf_pqfs(review_embedding)
INDEX_OPTIONS'{"index_type":"IVF_PQFS",
               "metric_type":"DOT_PRODUCT",
               "m":256,
               "nlist":1024,
               "nprobe":20}';

After the index is loaded, run the OPTIMIZE command for best performance.

OPTIMIZE TABLE reviews FULL;

Perform a Similarity Search Query

This example searches for the top 5 reviews that are most similar to a particular phrase. The query below first sets up a @query_vec, which is a vector that represents the phrase to be searched on. In the example below, the @query_vec is created from a JSON string; more commonly the @query_vec would be obtained from an API or from querying a different table. Refer to Working with Vector Data for more information on vector embeddings.

The query below uses an ORDER BY … LIMIT query to look for the top 5 reviews that are most similar to @query_vec. DESC ordering is used in the ORDER BY clause as the metric used to create the vector index was DOT_PRODUCT.

SET @query_vec=('[0.002, 0.01, 0.001, ...]'):>VECTOR(1024);

SELECT id, review,
   review_embedding <*> @query_vec AS score
FROM reviews
ORDER BY score USE INDEX (ivf_pqfs) SEARCH_OPTIONS '{"k":50}' DESC
LIMIT 5;

Key points to note about this query:

1. The @query_vec variable is cast to a VECTOR to ensure that @query_vec is a valid VECTOR and to improve performance.

2. The SEARCH_OPTIONS clause sets the search parameter k to 50. A setting of k equal to the the LIMIT value * 10 is a good starting point. In this example LIMIT is 5, so k is set to 5*10 = 50.

The USE INDEX clause in the query above specifies that the index named ivf_pqfs should be used in the query. USE INDEX is included for completeness, but is not necessary in this query. The index will be used regardless of whether the clause is included or not.

Verify a Vector Index is Being Used

It is important to verify that your query is using a vector index. A vector index can be used in the following conditions:

• The metric in the query (dot product vs. euclidean distance) matches the metric used to create the index.

• The order in the ORDER BY clause (DESC vs. ASC) in the search query matches the metric.

You can verify that a vector index is being used in a query with EXPLAIN as shown below.

SET @query_vec = ('[0.002, 0.01, 0.0001, ...]'):>VECTOR(1024);

EXPLAIN
SELECT id, review,
   review_embedding <*> @query_vec AS score
FROM reviews
ORDER BY score USE INDEX (ivf_pqfs) SEARCH_OPTIONS '{"k":50}' DESC
LIMIT 5;

+----------------------------------------------------------------------------------------------------------------------------+
| EXPLAIN                                                                                                                    |
+----------------------------------------------------------------------------------------------------------------------------+
| Project [remote_0.id, remote_0.review, remote_0.score]                                                                     |
| TopSort limit:5 [remote_0.score DESC]                                                                                      |
| Gather partitions:all alias:remote_0 parallelism_level:segment                                                             |
| Project [reviews.id, reviews.review, DOT_PRODUCT(reviews.review_embedding,(@query_vec:>vector(1024, F32) NOT NULL)) AS score] |
| TopSort limit:5 [DOT_PRODUCT(reviews.review_embedding,(@query_vec:>vector(3, F32) NOT NULL)) DESC]                         |
| ColumnStoreFilter [INTERNAL_VECTOR_SEARCH(0, (@query_vec:>vector(1024, F32) NOT NULL), 5, '{\"k\" : 50 }') index]             |
| ColumnStoreScan dbname.reviews, SORT KEY __UNORDERED () table_type:sharded_columnstore                                     |
+----------------------------------------------------------------------------------------------------------------------------+

The ColumnStoreFilter line of the EXPLAIN output above shows INTERNAL_VECTOR_SEARCH(...) which indicates that the vector index is being used. If INTERNAL_VECTOR_SEARCH is not in the ColumnStoreFilter line, the vector index is not being used.

Test Index Performance

To understand if an index meets the needs of your application, you will want to test the query accuracy (recall) and search time.

SET @query_vec = ('[0.002, 0.01, 0.0001, ...]'):>VECTOR(1024);

SELECT id, review,
   review_embedding <*> @query_vec AS score
FROM reviews
ORDER BY score USE INDEX() DESC
LIMIT 5;

SET @query_vec = ('[0.002, 0.01, 0.0001, ...]'):>VECTOR(1024);

PROFILE
SELECT id, review,
   review_embedding <*> @query_vec AS score
FROM reviews
ORDER BY score USE INDEX (ivf_pqfs) SEARCH_OPTIONS '{"k":50}' DESC
LIMIT 5;

m

The parameter m is an index building parameter that determines the number of subquantizers used in product quantization. Product quantization is used to compress high-dimensional vectors so they use less memory. The memory reduction in using product quantization is significant and is the reason that IVF_PQFS indexes use significantly less memory than HNSW_FLAT indexes.

The parameter m effectively trades off index size and accuracy of results. A higher m gives greater accuracy (recall) but the index size is larger. A lower value of m gives a smaller index, but the results will have lower accuracy. The parameter m must be set so that dimensions % m = 0 and m defaults to 32. Refer to Vector Indexing for details.

Consider values for m such as: dimensions/2, dimensions/4, dimensions/8. SingleStore recommends starting with dimensions/4.

k

The parameter k is an index search parameter which can be specified at query time. An example query using k is shown below.

The following query uses the reviews table defined above. As before, the @query_vec is a VECTOR representing a phrase the user is interested in and the query searches for reviews in the reviews table that are similar to @query_vec. Specifically, the query searches for the top five comments which are most similar to @query_vec.

SET @query_vec = ('[0.002, 0.01, 0.0001, ...]'):>VECTOR(1024);

SELECT id, review,
   review_embedding <*> @query_vec AS score
FROM reviews
ORDER BY score USE INDEX (ivf_pqfs) SEARCH_OPTIONS '{"k":50}' DESC
LIMIT 5;

In this query, k determines the number of rows output by the vector index scan, that is k determines the number of rows that the vector index scan produces before the ORDER BY … LIMIT clause is applied. Larger values of k give more accurate results (better recall), but increase search times. The parameter k must be >= limit, where limit is from ORDER BY … LIMIT clause and the default for k is the LIMIT value.

SingleStore suggests starting with a k of 10 * LIMIT and adjusting to get the recall and search time needed for your application.

nprobe

nprobe is a search parameter which specifies the number of probes to be used at query time. During index build, vectors are clustered into nlist clusters and then the query search is done by searching nprobe of the nearest clusters. SingleStore recommends setting nprobe to 20; nprobe cannot be greater than nlist.

A query using a vector index will search an nprobe/nlist fraction of the data. Increasing nprobe will improve accuracy at the cost of increased query time.

nlist

nlist is an index-building parameter which specifies the number of inverted lists created when the index is built. That is, nlist controls the number of centroids generated by k-means clustering. SingleStore recommends starting with nlist = 1024; nlist is the last of the parameters you should try tuning.

If you fix nprobe and increase nlist, you get faster queries with worse accuracy; however, the results depend on data distribution and may not change monotonically. A general recommendation is to set nlist to be of the same order as the square root of the number of rows in a segment.

The following query finds the segment size in number of rows for the reviews table.

SELECT database_name, table_name,
  AVG(rows_count) AS average_segment_size_in_rows
FROM information_schema.columnar_segments
WHERE database_name = 'dbname' AND table_name = 'reviews'
GROUP BY ALL;

1. Add a Vector Index

First, add a vector index on the table. This example uses a HNSW index of type HNSW_FLAT; create the type of index that you want to test.

ALTER TABLE reviews ADD VECTOR INDEX hnsw(review_embedding) 
   INDEX_OPTIONS '{"index_type":"HNSW_FLAT"}';

At this point the vector index has been built on disk, but has not been loaded into memory.

2. Identify the Current Amount of Memory used by Vector Indexes

Run SHOW STATUS EXTENDED LIKE 'alloc_vector_index' to get the amount of memory currently used by vector indexes.

3. Load the Vector Index into Memory

To force the vector index to be loaded into memory, run a query that uses the vector index.

To do so, run the following two queries which will cause the vector index to be loaded into memory. The first query selects a vector from the reviews table and the second query uses that vector in a nearest neighbor search.

In the first query, it is necessary to select a vector from the reviews table to use as a query vector in the subsequent query (as opposed to using a random query vector) in order to force the vector index to be loaded into memory.

SET @query_vec = 
(SELECT review_embedding 
           FROM reviews LIMIT 1);

SELECT @query_vec <*> review_embedding AS s 
INTO @s2
FROM reviews
ORDER BY s
USE INDEX (hnsw)
DESC LIMIT 1;

4. Find the Memory Use

Re-run SHOW STATUS EXTENDED LIKE 'alloc_vector_index' to get the amount of memory currently used by vector indexes. Take the difference between the value found in item 2. above and this value to obtain the amount of memory used by the vector index.