Haskell :: Aeson :: parse ADT based on field value

问题:

I'm using an external API which returns JSON responses. One of the responses is an array of objects and these objects are identified by the field value inside them. I'm having some trouble understanding how the parsing of such JSON response could be done with Aeson.

Here is a simplified version of my problem:

newtype Content = Content { content :: [Media] } deriving (Generic)

instance FromJSON Content

data Media =
  Video { objectClass :: Text
        , title :: Text } |
  AudioBook { objectClass :: Text
            , title :: Text }

In API documentation it is said that the object can be identified by the field objectClass which has value "video" for our Video object and "audiobook" for our AudioBook and so on. Example JSON:

[{objectClass: "video", title: "Some title"}
,{objectClass: "audiobook", title: "Other title"}]

The question is how can this type of JSON be approached using Aeson?

instance FromJSON Media where
  parseJSON (Object x) = ???

回答1:

You basically need a function Text -> Text -> Media:

toMedia :: Text -> Text -> Media
toMedia "video"     = Video "video"
toMedia "audiobook" = AudioBook "audiobook"

The FromJSON instance is now really simple (using <$> and <*> from Control.Applicative):

instance FromJSON Media where
    parseJSON (Object x) = toMedia <$> x .: "objectClass" <*> x .: "title"

However, at this point you're redundant: the objectClass field in Video or Audio doesn't give you more information than the actual type, so you might remove it:

data Media = Video     { title :: Text }
           | AudioBook { title :: Text }

toMedia :: Text -> Text -> Media
toMedia "video"     = Video
toMedia "audiobook" = AudioBook

Also note that toMedia is partial. You probably want to catch invalid "objectClass" values:

instance FromJSON Media where
    parseJSON (Object x) = 
        do oc <- x .: "objectClass"
           case oc of
               String "video"     -> Video     <$> x .: "title"
               String "audiobook" -> AudioBook <$> x .: "title"
               _                  -> empty

{- an alternative using a proper toMedia
toMedia :: Alternative f => Text -> f (Text -> Media)
toMedia "video"     = pure Video
toMedia "audiobook" = pure AudioBook
toMedia _           = empty

instance FromJSON Media where
    parseJSON (Object x) = (x .: "objectClass" >>= toMedia) <*> x .: "title"
-}

And last, but not least, remember that valid JSON uses strings for the name.

回答2:

The default translation for a data type like:

data Media = Video     { title :: Text }
           | AudioBook { title :: Text }
             deriving Generic

is actually very close to what you want. (For the simplicity of my examples, I define ToJSON instances and encode the examples to see what kind of JSON we get.)

aeson, default

So, with the default instance we have (view the complete source file which produces this output):

[{"tag":"Video","title":"Some title"},{"tag":"AudioBook","title":"Other title"}]

Let's see whether we can get even closer with custom options...

aeson, custom tagFieldName

With custom options:

mediaJSONOptions :: Options
mediaJSONOptions = 
    defaultOptions{ sumEncoding = 
                        TaggedObject{ tagFieldName = "objectClass"
                                    -- , contentsFieldName = undefined
                                    }
                  }

instance ToJSON Media
    where toJSON = genericToJSON mediaJSONOptions

we get:

[{"objectClass":"Video","title":"Some title"},{"objectClass":"AudioBook","title":"Other title"}]

(Think yourself what you want to do with an undefined field in the real code.)

aeson, custom constructorTagModifier

Adding

              , constructorTagModifier = fmap Char.toLower

to mediaJSONOptions gives:

[{"objectClass":"video","title":"Some title"},{"objectClass":"audiobook","title":"Other title"}]

Great! Exactly what you specified!

decoding

Simply add an instance with the same options to be able to decode from this format:

instance FromJSON Media
    where parseJSON = genericParseJSON mediaJSONOptions

Example:

*Main> encode example
"[{\"objectClass\":\"video\",\"title\":\"Some title\"},{\"objectClass\":\"audiobook\",\"title\":\"Other title\"}]"
*Main> decode $ fromString "[{\"objectClass\":\"video\",\"title\":\"Some title\"},{\"objectClass\":\"audiobook\",\"title\":\"Other title\"}]" :: Maybe [Media]
Just [Video {title = "Some title"},AudioBook {title = "Other title"}]
*Main>

Complete source file.

generic-aeson, default

To get a more complete picture, let's also look at what generic-aeson package would give (at hackage). It has also nice default translations, different in some respects from those from aeson.

Doing

import Generics.Generic.Aeson -- from generic-aeson package

and defining:

instance ToJSON Media
    where toJSON = gtoJson

gives the result:

[{"video":{"title":"Some title"}},{"audioBook":{"title":"Other title"}}]

So, it's different from all what we've seen when using aeson.

generic-aeson's options (Settings) are not interesting for us (they allow only to strip a prefix).

(The complete source file.)

aeson, ObjectWithSingleField

Apart from lower-casing the first letter of the constructor names, generic-aeson's translation seems similar to an option available in aeson:

Let's try this:

mediaJSONOptions = 
    defaultOptions{ sumEncoding = ObjectWithSingleField
                  , constructorTagModifier = fmap Char.toLower
                  }

and yes, the result is:

[{"video":{"title":"Some title"}},{"audiobook":{"title":"Other title"}}]

the rest of options: (aeson, TwoElemArray)

One available option for sumEncoding has been left out from consideration above, because it gives an array which is not quite similar to the JSON representation asked about. It's TwoElemArray. Example:

[["video",{"title":"Some title"}],["audiobook",{"title":"Other title"}]]

is given by:

mediaJSONOptions = 
    defaultOptions{ sumEncoding = TwoElemArray
                  , constructorTagModifier = fmap Char.toLower
                  }