7.6.2. Vector column

7.6.2.1. Summary

Vector column is a data store object. It can stores zero or more scalar values. In short, scalar value is a single value such as number and string. See Scalar column about scalar value details.

One of vector column use cases is tags store. You can use a vector column to store tag values.

You can use vector column as index search target in the same way as scalar column. You can set weight for each element. The element that has one or more weight is matched, the record has more score rather than no weight case. It is a vector column specific feature. Vector column that can store weight is called weight vector column.

You can also do full text search against each text element. But search score is too high when weight is used. You should use full text search with weight carefully.

7.6.2.2. Usage

There are three vector column types:

  • Normal vector column
  • Reference vector column
  • Weight vector column

This section describes how to use these types.

7.6.2.2.1. Normal vector column

Normal vector column stores zero or more scalar data. For example, scalar data are number, string and so on.

A normal vector column can store the same type elements. You can't mix types. For example, you can't store a number and a string in the same normal vector column.

Normal vector column is useful when a record has multiple values with a key. Tags are the most popular use case.

7.6.2.2.1.1. How to create

Use column_create command to create a normal vector column. The point is COLUMN_VECTOR flag:

Execution example:

table_create Bookmarks TABLE_HASH_KEY ShortText
# [[0, 1337566253.89858, 0.000355720520019531], true]
column_create Bookmarks tags COLUMN_VECTOR ShortText
# [[0, 1337566253.89858, 0.000355720520019531], true]

You can set zero or more tags to a bookmark.

7.6.2.2.1.2. How to load

You can load vector data by JSON array syntax:

[ELEMENT1, ELEMENT2, ELEMENT3, ...]

Let's load the following data:

_key tags
http://groonga.org/ ["groonga"]
http://mroonga.org/ ["mroonga", "mysql", "groonga"]
http://ranguba.org/ ["ruby", "groonga"]

Here is a command that loads the data:

Execution example:

load --table Bookmarks
[
{"_key": "http://groonga.org/", "tags": ["groonga"]},
{"_key": "http://mroonga.org/", "tags": ["mroonga", "mysql", "groonga"]},
{"_key": "http://ranguba.org/", "tags": ["ruby", "groonga"]}
]
# [[0, 1337566253.89858, 0.000355720520019531], 3]

The loaded data can be outputted as JSON array syntax:

Execution example:

select Bookmarks
# [
#   [
#     0,
#     1337566253.89858,
#     0.000355720520019531
#   ],
#   [
#     [
#       [
#         3
#       ],
#       [
#         [
#           "_id",
#           "UInt32"
#         ],
#         [
#           "_key",
#           "ShortText"
#         ],
#         [
#           "tags",
#           "ShortText"
#         ]
#       ],
#       [
#         1,
#         "http://groonga.org/",
#         [
#           "groonga"
#         ]
#       ],
#       [
#         2,
#         "http://mroonga.org/",
#         [
#           "mroonga",
#           "mysql",
#           "groonga"
#         ]
#       ],
#       [
#         3,
#         "http://ranguba.org/",
#         [
#           "ruby",
#           "groonga"
#         ]
#       ]
#     ]
#   ]
# ]

7.6.2.2.2. Reference vector column

TODO

Reference vector column is space-efficient if there are many same value elements. Reference vector column keeps reference record IDs not value itself. Record ID is smaller than value itself.

7.6.2.2.2.1. How to create

TODO

7.6.2.2.2.2. How to load

TODO

7.6.2.2.2.3. How to search

TODO

7.6.2.2.3. Weight vector column

Weight vector column is similar to normal vector column. It can store elements. It can also stores weights for them. Weight is degree of importance of the element.

Weight is positive integer. 0 is the default weight. It means that no weight.

If weight is one or larger, search score is increased by the weight. If the weight is 0, score is 1. If the weight is 10, score is 11 (= 1 + 10).

Weight vector column is useful for tuning search score. See also adjuster. You can increase search score of specific records.

7.6.2.2.3.1. Limitations

There are some limitations for now. They will be resolved in the future.

Here are limitations:

  • You need to use string representation for element value on load. For example, you can't use 29 for number 29. You need to use "29" for number 29.

7.6.2.2.3.2. How to create

Use column_create command to create a weight vector column. The point is COLUMN_VECTOR|WITH_WEIGHT flags:

Execution example:

table_create Bookmarks TABLE_HASH_KEY ShortText
# [[0, 1337566253.89858, 0.000355720520019531], true]
column_create Bookmarks tags COLUMN_VECTOR|WITH_WEIGHT ShortText
# [[0, 1337566253.89858, 0.000355720520019531], true]

If you don't specify WITH_WEIGHT flag, it is just a normal vector column.

You can set zero or more tags with weight to a bookmark.

7.6.2.2.3.3. How to load

You can load vector data by JSON object syntax:

{"ELEMENT1": WEIGHT1, "ELEMENT2": WEIGHT2, "ELEMENT3": WEIGHT3, ...}

Let's load the following data:

_key tags
http://groonga.org/ {"groonga": 100}
http://mroonga.org/ {"mroonga": 100, "mysql": 50, "groonga": 10}
http://ranguba.org/ {"ruby": 100, "groonga": 50}

Here is a command that loads the data:

Execution example:

load --table Bookmarks
[
{"_key": "http://groonga.org/",
 "tags": {"groonga": 100}},
{"_key": "http://mroonga.org/",
 "tags": {"mroonga": 100,
          "mysql":   50,
          "groonga": 10}},
{"_key": "http://ranguba.org/",
 "tags": {"ruby": 100,
          "groonga": 50}}
]
# [[0, 1337566253.89858, 0.000355720520019531], 3]

The loaded data can be outputted as JSON object syntax:

Execution example:

select Bookmarks
# [
#   [
#     0,
#     1337566253.89858,
#     0.000355720520019531
#   ],
#   [
#     [
#       [
#         3
#       ],
#       [
#         [
#           "_id",
#           "UInt32"
#         ],
#         [
#           "_key",
#           "ShortText"
#         ],
#         [
#           "tags",
#           "ShortText"
#         ]
#       ],
#       [
#         1,
#         "http://groonga.org/",
#         {
#           "groonga": 100
#         }
#       ],
#       [
#         2,
#         "http://mroonga.org/",
#         {
#           "mroonga": 100,
#           "groonga": 10,
#           "mysql": 50
#         }
#       ],
#       [
#         3,
#         "http://ranguba.org/",
#         {
#           "ruby": 100,
#           "groonga": 50
#         }
#       ]
#     ]
#   ]
# ]

7.6.2.2.3.4. How to search

You need to create an index to search weight vector column. You don't forget to specify WITH_WEIGHT flag to column_create:

Execution example:

table_create Tags TABLE_PAT_KEY ShortText
# [[0, 1337566253.89858, 0.000355720520019531], true]
column_create Tags bookmark_index COLUMN_INDEX|WITH_WEIGHT Bookmarks tags
# [[0, 1337566253.89858, 0.000355720520019531], true]

There is no weight vector column specific way except WITH_WEIGHT flag. You can create an index like a scalar column.

You can search an element in tags like full text search syntax.

With match_columns and query:

Execution example:

select Bookmarks --match_columns tags --query groonga --output_columns _key,tags,_score
# [
#   [
#     0,
#     1337566253.89858,
#     0.000355720520019531
#   ],
#   [
#     [
#       [
#         3
#       ],
#       [
#         [
#           "_key",
#           "ShortText"
#         ],
#         [
#           "tags",
#           "ShortText"
#         ],
#         [
#           "_score",
#           "Int32"
#         ]
#       ],
#       [
#         "http://groonga.org/",
#         {
#           "groonga": 100
#         },
#         101
#       ],
#       [
#         "http://mroonga.org/",
#         {
#           "mroonga": 100,
#           "groonga": 10,
#           "mysql": 50
#         },
#         11
#       ],
#       [
#         "http://ranguba.org/",
#         {
#           "ruby": 100,
#           "groonga": 50
#         },
#         51
#       ]
#     ]
#   ]
# ]

You can also use weight in match_columns. The score is (1 + weight_in_weight_vector) * weight_in_match_columns:

Execution example:

select Bookmarks --match_columns 'tags * 3' --query groonga --output_columns _key,tags,_score
# [
#   [
#     0,
#     1337566253.89858,
#     0.000355720520019531
#   ],
#   [
#     [
#       [
#         3
#       ],
#       [
#         [
#           "_key",
#           "ShortText"
#         ],
#         [
#           "tags",
#           "ShortText"
#         ],
#         [
#           "_score",
#           "Int32"
#         ]
#       ],
#       [
#         "http://groonga.org/",
#         {
#           "groonga": 100
#         },
#         303
#       ],
#       [
#         "http://mroonga.org/",
#         {
#           "mroonga": 100,
#           "groonga": 10,
#           "mysql": 50
#         },
#         33
#       ],
#       [
#         "http://ranguba.org/",
#         {
#           "ruby": 100,
#           "groonga": 50
#         },
#         153
#       ]
#     ]
#   ]
# ]

With filter:

Execution example:

select Bookmarks --filter 'tags @ "groonga"' --output_columns _key,tags,_score
# [
#   [
#     0,
#     1337566253.89858,
#     0.000355720520019531
#   ],
#   [
#     [
#       [
#         3
#       ],
#       [
#         [
#           "_key",
#           "ShortText"
#         ],
#         [
#           "tags",
#           "ShortText"
#         ],
#         [
#           "_score",
#           "Int32"
#         ]
#       ],
#       [
#         "http://groonga.org/",
#         {
#           "groonga": 100
#         },
#         101
#       ],
#       [
#         "http://mroonga.org/",
#         {
#           "mroonga": 100,
#           "groonga": 10,
#           "mysql": 50
#         },
#         11
#       ],
#       [
#         "http://ranguba.org/",
#         {
#           "ruby": 100,
#           "groonga": 50
#         },
#         51
#       ]
#     ]
#   ]
# ]

7.6.2.2.3.5. How to apply just weight

You can use weight in weight vector column to just increase search score without changing a set of matched records.

Use adjuster for the purpose:

Execution example:

select Bookmarks \
  --filter true \
  --adjuster 'tags @ "mysql" * 10 + tags @ "groonga" * 5' \
  --output_columns _key,tags,_score
# [
#   [
#     0,
#     1337566253.89858,
#     0.000355720520019531
#   ],
#   [
#     [
#       [
#         3
#       ],
#       [
#         [
#           "_key",
#           "ShortText"
#         ],
#         [
#           "tags",
#           "ShortText"
#         ],
#         [
#           "_score",
#           "Int32"
#         ]
#       ],
#       [
#         "http://groonga.org/",
#         {
#           "groonga": 100
#         },
#         506
#       ],
#       [
#         "http://mroonga.org/",
#         {
#           "mroonga": 100,
#           "groonga": 10,
#           "mysql": 50
#         },
#         566
#       ],
#       [
#         "http://ranguba.org/",
#         {
#           "ruby": 100,
#           "groonga": 50
#         },
#         256
#       ]
#     ]
#   ]
# ]

The select command uses --filter true. So all records are matched with score 1. Then it applies --adjuster. The adjuster does the following:

  • tags @ "mysql" * 10 increases score by (1 + weight) * 10 of records that has "mysql" tag.
  • tags @ "groonga" * 5 increases score by (1 + weight) * 5 of records that has "groonga" tag.

For example, record "http://mroonga.org/" has both "mysql" tag and "groonga" tag. So its score is increased by 565 (= ((1 + 50) * 10) + ((1 + 10) * 5) = (51 * 10) + (11 * 5) = 510 + 55). The search score is 1 by --filter true before applying --adjuster. So the final search score is 566 (= 1 + 565) of record "http://mroonga.org/".