Query cache stores compressed result sets in memory and reuses them for subsequent queries when possible. You can configure it using the following directives:

• qcache_max_bytes, a limit on the RAM usage for cached query storage. Defaults to 16 MB. Setting qcache_max_bytes to 0 completely disables the query cache.
• qcache_thresh_msec, the minimum wall query time to cache. Queries that complete faster than this will not be cached. Defaults to 3000 msec, or 3 seconds.
• qcache_ttl_sec, cached entry TTL, or time to live. Queries will stay cached for this duration. Defaults to 60 seconds, or 1 minute.

These settings can be changed on the fly using the SET GLOBAL statement:

mysql> SET GLOBAL qcache_max_bytes=128000000;

These changes are applied immediately, and cached result sets that no longer satisfy the constraints are immediately discarded. When reducing the cache size on the fly, MRU (most recently used) result sets win.

Query cache operates as follows. When enabled, every full-text search result is completely stored in memory. This occurs after full-text matching, filtering, and ranking, so essentially we store total_found {docid,weight} pairs. Compressed matches can consume anywhere from 2 bytes to 12 bytes per match on average, mostly depending on the deltas between subsequent docids. Once the query is complete, we check the wall time and size thresholds, and either save the compressed result set for reuse or discard it.

Note that the query cache's impact on RAM is not limited byqcache_max_bytes! If you run, for example, 10 concurrent queries, each matching up to 1M matches (after filters), then the peak temporary RAM usage will be in the range of 40 MB to 240 MB, even if the queries are fast enough and don't get cached.

Queries can use cache when the table, full-text query (i.e.,MATCH() contents), and ranker all match, and filters are compatible. This means:

• The full-text part within MATCH() must be a bytewise match. Add a single extra space, and it's now a different query as far as the query cache is concerned.
• The ranker (and its parameters, if any, for user-defined rankers) must be a bytewise match.
• The filters must be a superset of the original filters. You can add extra filters and still hit the cache. (In this case, the extra filters will be applied to the cached result.) But if you remove one, that will be a new query again.

Cache entries expire with TTL and also get invalidated on table rotation, or on TRUNCATE, or on ATTACH. Note that currently, entries are not invalidated on arbitrary RT table writes! So a cached query might return older results for the duration of its TTL.

You can inspect the current cache status with SHOW STATUS through the qcache_XXX variables:

mysql> SHOW STATUS LIKE 'qcache%';
+-----------------------+----------+
| Counter               | Value    |
+-----------------------+----------+
| qcache_max_bytes      | 16777216 |
| qcache_thresh_msec    | 3000     |
| qcache_ttl_sec        | 60       |
| qcache_cached_queries | 0        |
| qcache_used_bytes     | 0        |
| qcache_hits           | 0        |
+-----------------------+----------+
6 rows in set (0.00 sec)

Collations primarily impact string attribute comparisons. They define both the character set encoding and the strategy Manticore employs for comparing strings when performing ORDER BY or GROUP BY with a string attribute involved.

String attributes are stored as-is during indexing, and no character set or language information is attached to them. This is fine as long as Manticore only needs to store and return the strings to the calling application verbatim. However, when you ask Manticore to sort by a string value, the request immediately becomes ambiguous.

First, single-byte (ASCII, ISO-8859-1, or Windows-1251) strings need to be processed differently than UTF-8 strings, which may encode each character with a variable number of bytes. Thus, we need to know the character set type to properly interpret the raw bytes as meaningful characters.

Second, we also need to know the language-specific string sorting rules. For example, when sorting according to US rules in the en_US locale, the accented character ï (small letter i with diaeresis) should be placed somewhere after z. However, when sorting with French rules and the fr_FR locale in mind, it should be placed between i and j. Some other set of rules might choose to ignore accents altogether, allowing ï and i to be mixed arbitrarily.

Third, in some cases, we may require case-sensitive sorting, while in others, case-insensitive sorting is needed.

Collations encapsulate all of the following: the character set, the language rules, and the case sensitivity. Manticore currently provides four collations:

1. libc_ci
2. libc_cs
3. utf8_general_ci
4. binary

The first two collations rely on several standard C library (libc) calls and can thus support any locale installed on your system. They provide case-insensitive (_ci) and case-sensitive (_cs) comparisons, respectively. By default, they use the C locale, effectively resorting to bytewise comparisons. To change that, you need to specify a different available locale using the collation_libc_locale directive. The list of locales available on your system can usually be obtained with the locale command:

$ locale -a
C
en_AG
en_AU.utf8
en_BW.utf8
en_CA.utf8
en_DK.utf8
en_GB.utf8
en_HK.utf8
en_IE.utf8
en_IN
en_NG
en_NZ.utf8
en_PH.utf8
en_SG.utf8
en_US.utf8
en_ZA.utf8
en_ZW.utf8
es_ES
fr_FR
POSIX
ru_RU.utf8
ru_UA.utf8

The specific list of system locales may vary. Consult your OS documentation to install additional needed locales.

utf8_general_ci and binary locales are built-in into Manticore. The first one is a generic collation for UTF-8 data (without any so-called language tailoring); it should behave similarly to the utf8_general_ci collation in MySQL. The second one is a simple bytewise comparison.

Collation can be overridden via SQL on a per-session basis using the SET collation_connection statement. All subsequent SQL queries will use this collation. Otherwise, all queries will use the server default collation or as specified in the collation_server configuration directive. Manticore currently defaults to the libc_ci collation.

Collations affect all string attribute comparisons, including those within ORDER BY and GROUP BY, so differently ordered or grouped results can be returned depending on the collation chosen. Note that collations don't affect full-text searching; for that, use the charset_table.

When Manticore executes a fullscan query, it can either use a plain scan to check every document against the filters or employ additional data and/or algorithms to speed up query execution. Manticore uses a cost-based optimizer (CBO), also known as a "query optimizer" to determine which approach to take.

The CBO can also enhance the performance of full-text queries. See below for more details.

The CBO may decide to replace one or more query filters with one of the following entities if it determines that doing so will improve performance:

1. A docid index utilizes a special docid-only secondary index stored in files with the .spt extension. Besides improving filters on document IDs, the docid index is also used to accelerate document ID to row ID lookups and to speed up the application of large killlists during daemon startup.
2. A columnar scan relies on columnar storage and can only be used on a columnar attribute. It scans every value and tests it against the filter, but it is heavily optimized and is typically faster than the default approach.
3. Secondary indexes are generated for all attributes (except JSON) by default. They use the PGM index along with Manticore's built-in inverted index to retrieve the list of row IDs corresponding to a value or range of values. Secondary indexes are stored in files with the .spidx and .spjidx extensions. For information on how to generate secondary indexes over JSON attributes, see json_secondary_indexes.

The optimizer estimates the cost of each execution path using various attribute statistics, including:

1. Information on the data distribution within an attribute (histograms, stored in .sphi files). Histograms are generated automatically when data is indexed and serve as the primary source of information for the CBO.
2. Information from PGM (secondary indexes), which helps estimate the number of document lists to read. This assists in gauging doclist merge performance and in selecting the appropriate merge algorithm (priority queue merge or bitmap merge).
3. Columnar encoding statistics, employed to estimate columnar data decompression performance.
4. A columnar min-max tree. While the CBO uses histograms to estimate the number of documents left after applying the filter, it also needs to determine how many documents the filter had to process. For columnar attributes, partial evaluation of the min-max tree serves this purpose.
5. Full-text dictionary. The CBO utilizes term stats to estimate the cost of evaluating the full-text tree.

The optimizer computes the execution cost for every filter used in a query. Since certain filters can be replaced with several different entities (e.g., for a document id, Manticore can use a plain scan, a docid index lookup, a columnar scan (if the document id is columnar), and a secondary index), the optimizer evaluates all available combinations. However, there is a maximum limit of 1024 combinations.

To estimate query execution costs, the optimizer calculates the estimated costs of the most significant operations performed when executing the query. It uses preset constants to represent the cost of each operation.

The optimizer compares the costs of each execution path and chooses the path with the lowest cost to execute the query.

When working with full-text queries that have filters by attributes, the query optimizer decides between two possible execution paths. One is to execute the full-text query, retrieve the matches, and use filters. The other is to replace filters with one or more entities described above, fetch rowids from them, and inject them into the full-text matching tree. This way, full-text search results will intersect with full-scan results. The query optimizer estimates the cost of full-text tree evaluation and the best possible path for computing filter results. Using this information, the optimizer chooses the execution path.

Another factor to consider is multithreaded query execution (when pseudo_sharding is enabled). The CBO is aware that some queries can be executed in multiple threads and takes this into account. The CBO prioritizes shorter query execution times (i.e., latency) over throughput. For instance, if a query using a columnar scan can be executed in multiple threads (and occupy multiple CPU cores) and is faster than a query executed in a single thread using secondary indexes, multithreaded execution will be preferred.

Queries using secondary indexes and docid indexes always run in a single thread, as benchmarks indicate that there is little to no benefit in making them multithreaded.

At present, the optimizer only uses CPU costs and does not take memory or disk usage into account.

Manticore Search supports the ability to add embeddings generated by your Machine Learning models to each document, and then doing a nearest-neighbor search on them. This lets you build features like similarity search, recommendations, semantic search, and relevance ranking based on NLP algorithms, among others, including image, video, and sound searches.

An embedding is a method of representing data—such as text, images, or sound—as vectors in a high-dimensional space. These vectors are crafted to ensure that the distance between them reflects the similarity of the data they represent. This process typically employs algorithms like word embeddings (e.g., Word2Vec, BERT) for text or neural networks for images. The high-dimensional nature of the vector space, with many components per vector, allows for the representation of complex and nuanced relationships between items. Their similarity is gauged by the distance between these vectors, often measured using methods like Euclidean distance or cosine similarity.

Manticore Search enables k-nearest neighbor (KNN) vector searches using the HNSW library. This functionality is part of the Manticore Columnar Library.

create table test ( title text, image_vector float_vector knn_type='hnsw' knn_dims='4' hnsw_similarity='l2' );

Query OK, 0 rows affected (0.01 sec)

insert into test values ( 1, 'yellow bag', (0.653448,0.192478,0.017971,0.339821) ), ( 2, 'white bag', (-0.148894,0.748278,0.091892,-0.095406) );

POST /insert
{
    "table":"test_vec",
    "id":1,
    "doc":  { "title" : "yellow bag", "image_vector" : [0.653448,0.192478,0.017971,0.339821] }
}

POST /insert
{
    "table":"test_vec",
    "id":2,
    "doc":  { "title" : "white bag", "image_vector" : [-0.148894,0.748278,0.091892,-0.095406] }
}

Query OK, 2 rows affected (0.00 sec)

{
    "table":"test",
    "_id":1,
    "created":true,
    "result":"created",
    "status":201
}

{
    "table":"test",
    "_id":2,
    "created":true,
    "result":"created",
    "status":201
}

select id, knn_dist() from test where knn ( image_vector, 5, (0.286569,-0.031816,0.066684,0.032926), 2000 );

POST /search
{
    "table": "test",
    "knn":
    {
        "field": "image_vector",
        "query_vector": [0.286569,-0.031816,0.066684,0.032926],
        "k": 5,
        "ef": 2000
    }
}

+------+------------+
| id   | knn_dist() |
+------+------------+
|    1 | 0.28146550 |
|    2 | 0.81527930 |
+------+------------+
2 rows in set (0.00 sec)

{
    "took":0,
    "timed_out":false,
    "hits":
    {
        "total":2,
        "total_relation":"eq",
        "hits":
        [
            {
                "_id": 1,
                "_score":1,
                "_knn_dist":0.28146550,
                "_source":
                {
                    "title":"yellow bag",
                    "image_vector":[0.653448,0.192478,0.017971,0.339821]
                }
            },
            {
                "_id": 2,
                "_score":1,
                "_knn_dist":0.81527930,
                "_source":
                {
                    "title":"white bag",
                    "image_vector":[-0.148894,0.748278,0.091892,-0.095406]
                }
            }
        ]
    }
}

select id, knn_dist() from test where knn ( image_vector, 5, 1 );

POST /search
{
  "table": "test",
  "knn":
  {
    "field": "image_vector",
    "doc_id": 1,
    "k": 5
  }
}

+------+------------+
| id   | knn_dist() |
+------+------------+
|    2 | 0.81527930 |
+------+------------+
1 row in set (0.00 sec)

{
    "took":0,
    "timed_out":false,
    "hits":
    {
        "total":1,
        "total_relation":"eq",
        "hits":
        [
            {
                "_id": 2,
                "_score":1643,
                "_knn_dist":0.81527930,
                "_source":
                {
                    "title":"white bag",
                    "image_vector":[-0.148894,0.748278,0.091892,-0.095406]
                }
            }
        ]
    }
}

select id, knn_dist() from test where knn ( image_vector, 5, (0.286569,-0.031816,0.066684,0.032926) ) and match('white') and id < 10;

POST /search
{
    "table": "test",
    "knn":
    {
        "field": "image_vector",
        "query_vector": [0.286569,-0.031816,0.066684,0.032926],
        "k": 5,
        "filter":
        {
            "bool":
            {
                "must":
                [
                    { "match": {"_all":"white"} },
                    { "range": { "id": { "lt": 10 } } }
                ]
            }
        }
    }
}

+------+------------+
| id   | knn_dist() |
+------+------------+
|    2 | 0.81527930 |
+------+------------+
1 row in set (0.00 sec)

{
    "took":0,
    "timed_out":false,
    "hits":
    {
        "total":1,
        "total_relation":"eq",
        "hits":
        [
            {
                "_id": 2,
                "_score":1643,
                "_knn_dist":0.81527930,
                "_source":
                {
                    "title":"white bag",
                    "image_vector":[-0.148894,0.748278,0.091892,-0.095406]
                }
            }
        ]
    }
}