JSON Type Provider

This article demonstrates how to use the JSON Type Provider to access JSON files in a statically typed way. We first look at how the structure is inferred and then demonstrate the provider by parsing data from WorldBank and Twitter.

The JSON Type Provider provides statically typed access to JSON documents. It takes a sample document as an input (or a document containing a JSON array of samples). The generated type can then be used to read files with the same structure.

If the loaded file does not match the structure of the sample, a runtime error may occur (but only when explicitly accessing an element incompatible with the original sample — e.g. if it is no longer present).

Introducing the provider

The type provider is located in the FSharp.Data.dll assembly and namespace:

open FSharp.Data

Inferring types from the sample

The JsonProvider<...> takes one static parameter of type string. The parameter can be either a sample string or a sample file (relative to the current folder or online accessible via http or https). It is not likely that this could lead to ambiguities.

The following sample passes a small JSON string to the provider:

type Simple = JsonProvider<""" { "name":"John", "age":94 } """>
let simple = Simple.Parse(""" { "name":"Tomas", "age":4 } """)
simple.Age
simple.Name

type Simple = JsonProvider<...>
val simple: JsonProvider<...>.Root = {
  "name": "Tomas",
  "age": 4
}
val it: string = "Tomas"

You can see that the generated type has two properties - Age of type int and Name of type string. The provider successfully infers the types from the sample and exposes the fields as properties (with PascalCase name to follow standard naming conventions).

Inferring numeric types

In the previous case, the sample document simply contained an integer and so the provider inferred the type int. Sometimes, the types in the sample document (or a list of samples) may not match exactly. For example, a list may mix integers and floats:

type Numbers = JsonProvider<""" [1, 2, 3, 3.14] """>
let nums = Numbers.Parse(""" [1.2, 45.1, 98.2, 5] """)
let total = nums |> Seq.sum

type Numbers = JsonProvider<...>
val nums: decimal array = [|1.2M; 45.1M; 98.2M; 5M|]
val total: decimal = 149.5M

When the sample is a collection, the type provider generates a type that can be used to store all values in the sample. In this case, the resulting type is decimal, because one of the values is not an integer. In general, the provider supports (and prefers them in this order): int, int64, decimal and float.

Other primitive types cannot be combined into a single type. For example, if the list contains numbers and strings. In this case, the provider generates two methods that can be used to get values that match one of the types:

type Mixed = JsonProvider<""" [1, 2, "hello", "world"] """>
let mixed = Mixed.Parse(""" [4, 5, "hello", "world" ] """)

mixed.Numbers |> Seq.sum
mixed.Strings |> String.concat ", "

type Mixed = JsonProvider<...>
val mixed: JsonProvider<...>.Root = [
  4,
  5,
  "hello",
  "world"
]
val it: string = "4, 5, hello, world"

As you can see, the Mixed type has properties Numbers and Strings that return only int and string values from the collection. This means that we get type-safe access to the values, but not in the original order (if order matters, then you can use the mixed.JsonValue property to get the underlying JsonValue and process it dynamically as described in the documentation for JsonValue.

Inferring record types

Now let's look at a sample JSON document that contains a list of records. The following example uses two records - one with name and age and the second with just name. If a property is missing, then the provider infers it as optional.

If we want to just use the same text used for the schema at runtime, we can use the GetSamples method:

type People =
    JsonProvider<"""
  [ { "name":"John", "age":94 },
    { "name":"Tomas" } ] """>

for item in People.GetSamples() do
    printf "%s " item.Name
    item.Age |> Option.iter (printf "(%d)")
    printfn ""

John (94)
Tomas 
type People = JsonProvider<...>
val it: unit = ()

The inferred type for items is a collection of (anonymous) JSON entities - each entity has properties Name and Age. As Age is not available for all records in the sample data set, it is inferred as option<int>. The above sample uses Option.iter to print the value only when it is available.

In the previous case, the values of individual properties had common types - string for the Name property and numeric type for Age. However, what if the property of a record can have multiple different types? In that case, the type provider behaves as follows:

type Values = JsonProvider<""" [{"value":94 }, {"value":"Tomas" }] """>

for item in Values.GetSamples() do
    match item.Value.Number, item.Value.String with
    | Some num, _ -> printfn "Numeric: %d" num
    | _, Some str -> printfn "Text: %s" str
    | _ -> printfn "Some other value!"

Numeric: 94
Text: Tomas
type Values = JsonProvider<...>
val it: unit = ()

Here, the Value property is either a number or a string, The type provider generates a type that has an optional property for each possible option, so we can use simple pattern matching on option<int> and option<string> values to distinguish between the two options. This is similar to the handling of heterogeneous arrays.

Note that we have a GetSamples method because the sample is a JSON list. If it was a JSON object, we would have a GetSample method instead.

More complex object type on root level

If you want the root type to be an object type, not an array, but you need more samples at root level, you can use the SampleIsList parameter. Applied to the previous example this would be:

type People2 =
    JsonProvider<"""
  [ { "name":"John", "age":94 },
    { "name":"Tomas" } ] """, SampleIsList=true>

let person = People2.Parse("""{ "name":"Gustavo" }""")

type People2 = JsonProvider<...>
val person: JsonProvider<...>.Root = {
  "name": "Gustavo"
}

Note that starting with version 4.2.9 of this package, JSON comments are supported (Comments are either single-line and start with // or multi-line when wrapped in /* and */). This is not a standard feature of JSON, but it can be really convenient, e.g. to annotate each sample when using multiple ones.

Type inference hints / inline schemas

Starting with version 4.2.10 of this package, it's possible to enable basic type annotations directly in the sample used by the provider, to complete or to override type inference. (Only basic types are supported. See the reference documentation of the provider for the full list)

This feature is disabled by default and has to be explicitly enabled with the InferenceMode static parameter.

Let's consider an example where this can be useful:

type AmbiguousEntity =
    JsonProvider<Sample="""
        { "code":"000", "length":"0" }
        { "code":"123", "length":"42" }
        { "code":"4E5", "length":"1.83" }
        """, SampleIsList=true>

let code = (AmbiguousEntity.GetSamples()[1]).Code
let length = (AmbiguousEntity.GetSamples()[1]).Length

type AmbiguousEntity = JsonProvider<...>
val code: float = 123.0
val length: decimal = 42M

In the previous example, Code is inferred as a float, even though it looks more like it should be a string. (4E5 is interpreted as an exponential float notation instead of a string)

Now let's enable inline schemas:

open FSharp.Data.Runtime.StructuralInference

type AmbiguousEntity2 =
    JsonProvider<Sample="""
        { "code":"typeof<string>", "length":"typeof< float<metre> >" }
        { "code":"123", "length":"42" }
        { "code":"4E5", "length":"1.83" }
        """, SampleIsList=true, InferenceMode=InferenceMode.ValuesAndInlineSchemasOverrides>

let code2 = (AmbiguousEntity2.GetSamples().[1]).Code
let length2 = (AmbiguousEntity2.GetSamples().[1]).Length

type AmbiguousEntity2 = JsonProvider<...>
val code2: string = "123"
val length2: JsonProvider<...>.DecimalOrString = "42"

With the ValuesAndInlineSchemasOverrides inference mode, the typeof<string> inline schema takes priority over the type inferred from other values. Code is now a string, as we wanted it to be!

Note that an alternative to obtain the same result would have been to replace all the Code values in the samples with unambiguous string values. (But this can be very cumbersome, especially with big samples)

If we had used the ValuesAndInlineSchemasHints inference mode instead, our inline schema would have had the same precedence as the types inferred from other values, and Code would have been inferred as a choice between either a number or a string, exactly as if we had added another sample with an unambiguous string value for Code.

You can use either angle brackets <> or curly brackets {} when defining inline schemas.

Units of measure

Inline schemas also enable support for units of measure.

In the previous example, the Length property is now inferred as a float with the metre unit of measure (from the default SI units).

Warning: units of measures are discarded when merged with types without a unit or with a different unit. As mentioned previously, with the ValuesAndInlineSchemasHints inference mode, inline schemas types are merged with other inferred types with the same precedence. Since values-inferred types never have units, inline-schemas-inferred types will lose their unit if the sample contains other values...

Loading WorldBank data

Now let's use the type provider to process some real data. We use a data set returned by the WorldBank, which has (roughly) the following structure:

[ { "page": 1, "pages": 1, "total": 53 },
  [ { "indicator": {"value": "Central government debt, total (% of GDP)"},
      "country": {"id":"CZ","value":"Czech Republic"},
      "value":null,"decimal":"1","date":"2000"},
    { "indicator": {"value": "Central government debt, total (% of GDP)"},
      "country": {"id":"CZ","value":"Czech Republic"},
      "value":"16.6567773464055","decimal":"1","date":"2010"} ] ]

The response to a request contains an array with two items. The first item is a record with general information about the response (page, total pages, etc.) and the second item is another array which contains the actual data points. For every data point, we get some information and the actual value. Note that the value is passed as a string (for some unknown reason). It is wrapped in quotes, so the provider infers its type as string (and we need to convert it manually).

The following sample generates type based on the data/WorldBank.json file and loads it:

[<Literal>]
let ResolutionFolder = __SOURCE_DIRECTORY__

type WorldBank = JsonProvider<"../data/WorldBank.json", ResolutionFolder=ResolutionFolder>
let doc = WorldBank.GetSample()

Note that we can also load the data directly from the web both in the Load method and in the type provider sample parameter, and there's an asynchronous AsyncLoad method available too:

let wbReq =
    "http://api.worldbank.org/country/cz/indicator/"
    + "GC.DOD.TOTL.GD.ZS?format=json"

let docAsync = WorldBank.AsyncLoad(wbReq)

val wbReq: string =
  "http://api.worldbank.org/country/cz/indicator/GC.DOD.TOTL.GD."+[14 chars]
val docAsync: Async<JsonProvider<...>.Root>

The doc is an array of heterogeneous types, so the provider generates a type that can be used to get the record and the array, respectively. Note that the provider infers that there is only one record and one array. We can print the data set as follows:

// Print general information
let info = doc.Record
printfn "Showing page %d of %d. Total records %d" info.Page info.Pages info.Total

// Print all data points
for record in doc.Array do
    record.Value
    |> Option.iter (fun value -> printfn "%d: %f" record.Date value)

Showing page 1 of 1. Total records 53
2010: 35.142297
2009: 31.034880
2008: 25.475164
2007: 24.193320
2006: 23.708055
2005: 22.033462
2004: 20.108379
2003: 18.267725
2002: 15.425565
2001: 14.874434
2000: 13.218869
1999: 11.356696
1998: 10.178780
1997: 10.153566
1996: 10.520301
1995: 12.707834
1994: 14.781808
1993: 16.656777
val info: JsonProvider<...>.Record2 =
  {
  "page": 1,
  "pages": 1,
  "per_page": "1000",
  "total": 53
}
val it: unit = ()

When printing the data points, some of the values might be missing (in the input, the value is null instead of a valid number). This is another example of a heterogeneous type - the type is either Number or some other type (representing null value). This means that record.Value has a Number property (when the value is a number) and we can use it to print the result only when the data point is available.

Parsing Twitter stream

We now look on how to parse tweets returned by the Twitter API. Tweets are quite heterogeneous, so we infer the structure from a list of inputs rather than from just a single input. To do that, we use the file data/TwitterStream.json (containing a list of tweets) and pass an optional parameter SampleIsList=true which tells the provider that the sample is actually a list of samples:

type Tweet = JsonProvider<"../data/TwitterStream.json", SampleIsList=true, ResolutionFolder=ResolutionFolder>

let text = (omitted)

let tweet = Tweet.Parse(text)

printfn "%s (retweeted %d times)\n:%s" tweet.User.Value.Name tweet.RetweetCount.Value tweet.Text.Value

After creating the Tweet type, we parse a single sample tweet and print some details about the tweet. As you can see, the tweet.User property has been inferred as optional (meaning that a tweet might not have an author?) so we unsafely get the value using the Value property. The RetweetCount and Text properties may be also missing, so we also access them unsafely.

Getting and creating GitHub issues

In this example we will now also create JSON in addition to consuming it. Let's start by listing the 5 most recently updated open issues in the FSharp.Data repository.

// GitHub.json downloaded from
// https://api.github.com/repos/fsharp/FSharp.Data/issues
// to prevent rate limit when generating these docs
type GitHub = JsonProvider<"../data/GitHub.json", ResolutionFolder=ResolutionFolder>

let topRecentlyUpdatedIssues =
    GitHub.GetSamples()
    |> Seq.filter (fun issue -> issue.State = "open")
    |> Seq.sortBy (fun issue -> System.DateTimeOffset.Now - issue.UpdatedAt)
    |> Seq.truncate 5

for issue in topRecentlyUpdatedIssues do
    printfn "#%d %s" issue.Number issue.Title

#879 Bug when call request from Http module
#867 XmlProvider in 2.2.5 on F# 4 project causes multiple FSharp.Core assembly references
#877 Header being considered as data row in HTMLProvider
#878 Replace GitHub JsonProvider example with something else in ConsoleTests because of rate limit
#873 Fix HtmlInference inferListType when passing an empty seq
type GitHub = JsonProvider<...>
val topRecentlyUpdatedIssues: JsonProvider<...>.Root seq
val it: unit = ()

And now let's create a new issue. We look into the documentation at http://developer.github.com/v3/issues/#create-an-issue and we see that we need to post a JSON value similar to this:

[<Literal>]
let issueSample =
    """
{
  "title": "Found a bug",
  "body": "I'm having a problem with this.",
  "assignee": "octocat",
  "milestone": 1,
  "labels": [
    "Label1",
    "Label2"
  ]
}
"""

This JSON is different from what we got for each issue in the previous API call, so we'll define a new type based on this sample, create an instance, and send a POST request:

type GitHubIssue = JsonProvider<issueSample, RootName="issue">

let newIssue =
    GitHubIssue.Issue(
        "Test issue",
        "This is a test issue created in FSharp.Data documentation",
        assignee = "",
        labels = [||],
        milestone = 0
    )

newIssue.JsonValue.Request "https://api.github.com/repos/fsharp/FSharp.Data/issues"

Using JSON provider in a library

You can use the types created by JSON type provider in a public API of a library that you are building, but there is one important thing to keep in mind - when the user references your library, the type provider will be loaded and the types will be generated at that time (the JSON provider is not currently a generative type provider). This means that the type provider will need to be able to access the sample JSON. This works fine when the sample is specified inline, but it won't work when the sample is specified as a local file (unless you distribute the samples with your library).

For this reason, the JSON provider lets you specify samples as embedded resources using the static parameter EmbeddedResource (don't forget then to include the file as EmbeddedResource in the project file). If you are building a library MyLib.dll, you can write:

type WB =
    JsonProvider<"../data/WorldBank.json", EmbeddedResource="MyLib, MyLib.data.worldbank.json", ResolutionFolder=ResolutionFolder>

You still need to specify the local path, but this is only used when compiling MyLib.dll. When a user of your library references MyLib.dll later, the JSON Type Provider will be able to load MyLib.dll and locate the sample worldbank.json as a resource of the library. When this succeeds, it does not attempt to find the local file and so your library can be used without providing a local copy of the sample JSON files.

Related articles

• JSON Parser - provides more information about working with JSON values dynamically.
• API Reference: JsonProvider
• API Reference: JsonValue