This section describes about the following completion features:
The completion feature uses three searches to compute completed words:
- Prefix RK search against registered words.
- Cooccurrence search against learned data.
- Prefix search against registered words. (optional)
RK means Romaji and Katakana. Prefix RK search can find registered words that start with user's input by romaji, katakana or hiragana. It's useful for searching in Japanese.
For example, there is a registered word "日本". And "ニホン" (it must be katakana) is registered as its reading. An user can find "日本" by "ni", "二" or "に".
Cooccurrence search can find registered words from user's partial input. It uses user input sequences that will be learned from query logs, access logs and so on.
For example, there is the following user input sequence:
| input | submit |
|---|---|
| s | no |
| se | no |
| sea | no |
| sear | no |
| searc | no |
| search | yes |
| e | no |
| en | no |
| eng | no |
| engi | no |
| engin | no |
| engine | no |
| enginen | no (typo!) |
| engine | yes |
Groonga creates the following completion pairs:
| input | completed word |
|---|---|
| s | search |
| se | search |
| sea | search |
| sear | search |
| searc | search |
| e | engine |
| en | engine |
| eng | engine |
| engi | engine |
| engin | engine |
| engine | engine |
| enginen | engine |
All user not-submitted inputs (e.g. "s", "se" and so on) before each an user submission maps to the submitted input (e.g. "search").
To be precise, this description isn't correct because it omits about time stamp. Groonga doesn't case about "all user not-submitted inputs before each an user submission". Groonga just case about "all user not-submitted inputs within a minute from an user submission before each an user submission". Groonga doesn't treat user inputs before a minute ago.
If an user inputs "sea" and cooccurrence search returns "search" because "sea" is in input column and corresponding completed word column value is "search".
Prefix search can find registered word that start with user's input. This search doesn't care about romaji, katakana and hiragana not like Prefix RK search.
This search isn't always ran. It's just ran when it's requested explicitly or both prefix RK search and cooccurrence search return nothing.
For example, there is a registered word "search". An user can find "search" by "s", "se", "sea", "sear", "searc" and "search".
Groonga provides suggest command to use completion. --type complete option requests completion.
For example, here is an command to get completion results by "en":
Execution example:
日本語版はこちらsuggest --table item_query --column kana --types complete --frequency_threshold 1 --query en
# [
# [
# 0,
# 1337566253.89858,
# 0.000355720520019531
# ],
# {
# "complete": [
# [
# 1
# ],
# [
# [
# "_key",
# "ShortText"
# ],
# [
# "_score",
# "Int32"
# ]
# ],
# [
# "engine",
# 1
# ]
# ]
# }
# ]
Cooccurrence search uses learned data. They are based on query logs, access logs and so on. To create learned data, groonga needs user input sequence with time stamp and user submit input with time stamp.
For example, an user wants to search by "engine". The user inputs the query with the following sequence:
- 2011-08-10T13:33:23+09:00: e
- 2011-08-10T13:33:23+09:00: en
- 2011-08-10T13:33:24+09:00: eng
- 2011-08-10T13:33:24+09:00: engi
- 2011-08-10T13:33:24+09:00: engin
- 2011-08-10T13:33:25+09:00: engine (submit!)
Groonga can be learned from the input sequence by the following command:
load --table event_query --each 'suggest_preparer(_id, type, item, sequence, time, pair_query)'
[
{"sequence": "1", "time": 1312950803.86057, "item": "e"},
{"sequence": "1", "time": 1312950803.96857, "item": "en"},
{"sequence": "1", "time": 1312950804.26057, "item": "eng"},
{"sequence": "1", "time": 1312950804.56057, "item": "engi"},
{"sequence": "1", "time": 1312950804.76057, "item": "engin"},
{"sequence": "1", "time": 1312950805.86057, "item": "engine", "type": "submit"}
]