These endpoints provide word predictions and auto-corrections based on client-provided distributions over possible characters.

Here are some test pages for the endpoints:

Distribution decoding tests

rec/distrib

Returns the most likely text given a sequence of distributions over possible characters. Optionally can return word completion predictions based on the observation sequence being the prefix of a word. Call should be via a POST containing JSON with the following key/value pairs:

Parameter	Required	Description	Valid values	Default
left	no	Left text context	One or more words separated by spaces	(none)
right	no	Right text context	One or more words separated by spaces	(none)
numBest	no	Number of best recognition results	Positive integer	5
numPrefix	no	Number of prefix word completion results	Non-negative integer	0
vocab	no	Filter to text only containing words in the specified vocabulary	1k, 5k, 10k, 20k, 40k, 100k, 500k	100k
sort	no	Sorting of result words	alpha = alphabetically A-Z rev-alpha = reverse alphabetically Z-A logprob = log probability, largest first rev-logprob = log probability, smallest first	logprob
safe	no	Safe mode, remove hypotheses that contain any obscene words	true or false	false
singleWord	no	Restricts recognition hypotheses to single words	true or false	true
exactContext	no	Client is providing exact left and right context including any spaces or sentence start/end words (for experts)	true or false	false
lang	no	Language code	Language to use (see list)	en
config	no	Configuration code	Recognition configuration to use (see list)	en
distribs	yes	List of distributions to recognize against	See details below	(none)

The distribs parameter is a list of the user's input distributions. Each distribution contains the following:

Key	Required	Description	Valid values	Default
distrib	yes	Distributions over possible characters	A list with 0 or more items containing: `text` and `logProb`.	(none)
time	no	Time of input	double value	(none)

Notes:

Best recognition hypotheses assume the tap sequence represents all the user's desired word(s).
Prefix recognition hypotheses assume the end of the tap sequence may only be the prefix of the desired word.
If list is empty and numPrefix > 0, returns the most probable words given surrounding text.
If singleWord is true, input is assumed to potentially contain observations from multiple words with or without spaces between them.
time is the time of the observation in seconds from some arbitrary starting reference point (e.g. the start of input).
logProb is assumed to be in base e.

Example POST request:

BASE_URL/rec/disitrib

{
    "left"       : "this is a",
    "right"      : "message",
    "numBest"    : 2,
    "numPrefix"  : 3,
    "vocab"      : "100k",
    "sort"       : "logprob",
    "safe"       : false,
    "singleWord" : true,
    "distribs"   : [ { "distrib" : [ {"text": "t", "logProb": -1.23},
                                     {"text": "r", "logProb": -2.34},
                                     {"text": "y", "logProb": -3.45}
                                   ],
                       "time"    : 0.123
                     },
                     { "distrib" : [ {"text": "e", "logProb": -1.56},
                                     {"text": "w", "logProb": -3.43},
                                     {"text": "r", "logProb": -4.84}
                                   ],
                       "time"    : 0.234
                     }
                   ]
}

Example response:

{
    "best": [
        {
            "text": "te",
            "logProb": -14.897991
        },
        {
            "text": "re",
            "logProb": -16.703276
        }
    ],
    "prefix": [
        {
            "text": "text",
            "logProb": -10.180267
        },
        {
            "text": "team",
            "logProb": -12.689505
        },
        {
            "text": "terror",
            "logProb": -13.126982
        }
    ]
}

document.write(PAGE_TITLE);

rec/distrib