Skip to content
Switch branches/tags
Go to file

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


An exercise of Web tools for supporting the embedding of Binary data in JSON responses with the following objectives:

  • avoid base64 bandwidth and decoding overhead
  • avoid custom encodings such MessagePack that are slow in browser decoding
  • reduce number of requests or additional production logic (if not using HTTP 2.0 Server Push)
  • can be used also for offline storage

Usage scenario: rich content produced on the fly from an embedded web server.


Offline Example without server: ([].

In this case the JSON-bag is baked into the HTML and the XHR is simulated. The output shows two images decoded from JSON and embedded. It has been tested with Chrome 55, Safari 10.1 and Firefox 52.

Online Example with the server: build with CMake the web server and run it passing the path of the root, then connect to

mkdir build
cd build
cmake ..
./example ..

Online Example interface:

  • /json returns the base64 version
  • /jsonbin returns the JSON-bag
  • /jsonlocal saves output.json with base64
  • /jsonbaglocal saves the JSON-bag binary


Create an JSON-bag response the acts much like a multipart MIME with binary data. We would like to have a solution that is human-readable (for the JSON part) and then directly addressable in the browser.

Having found that Binary glTF as KHR_binary_glTF addresses a similar problem JSON-bag provides a general solution.

Apart the Binary glTF specific domain the differences are:

  • patching of the JSON for supporting resource access
  • readable JSON heading

Format Diagream

HTTP Transfer


  • header
  • JSON
  • crlf null
  • blob1: header + data
  • blob2: header + data
  • ...

The header at the beginning tells the length of the JSON part, while the additional BLOB headers contain:

  • size
  • content-type
  • JSON pointer for patching the JSON

To simplify the decoding the BLOB header is in JSON prefixed with a 16-bit length

The first header could be conceived in a way to minimize disruption of content access by legacy services or diagnostics. Three possible approaches have been considered:

  1. textual header with magic, size in hexadecimal and crlf. Fixed length or variable
  2. pass the size in the HTTP header
  3. encapsulate the original JSON in a known-size JSON like "[size,original]"

Some comments on the approaches: (1) requires to skip the first line, and then JSON follows, being textual it is easy, (2) can be combined with the previous one, not suitable for file serialization, (3) a reasonable trick.

Note: Binary glTF enforces alignment of the BLOB payload


In JSON-bag different URL types are used in the JSON depends the transfer:

  • data: the classic base64 embedding, the simplest scenario
  • http: storing the contents in separate URL
  • blob: the new URL type supported by the Blob class and created as a view over the ArrayBuffer
  • refdata: a custom URL that is used to mark the offset, size and mime of an entity wrt the binary part

File Serialization

The JSON-bag could be stored as a JSON file plus one or more binary files (as glTF 2.0 does) or a single file. In this case the first header has to be stored in the file.

Current Format

The JSON bag format is the following usign the ABNF notation:

jsonbag = header JSONcontent separator *blob
header = magic version headersize crlf
headersize = hex sie
magic = "JBAG"
version = "0000"
separator = crlf \x00
blob = blob-header blob-payload
blob-header = headersize JSONcontent
headersize = uint16-le
blob-payload = *uin8

In the JSON expressing the blob header the following is necessary:

  • src (string)
  • mime (string)
  • size (integer)

Browser Side

The requests pulls AJAX data with special Accept header. Then receives the response and decodes first the JSON then uses the recently added Blob objects for creating views of data over the ArrayBuffer

var blob = new Blob( [ arrayBufferView ], { type: "image/jpeg" } );
var urlCreator = window.URL || window.webkitURL;
image.src = urlCreator.createObjectURL( blob );

The browser does the following:

  • read the JSON as part of the array buffer received using the size
  • read the declarations of binary parts, creates Blob object, obtain the URL and then patches the JSON in the correct path

Server Side

Example in C++ using Mongoose. The C++ API is independent of the JSON-bag and can produce the single JSON-bag or a JSON with multiple URLs. If the Accept header is present it serves the JSON-bag otherwise the server can decide between split URL version or base64 embedding. In anycase the URL in the JSON content is patched.

In the split version it patches the JSON with the effective URL, while in the single version it puts a placeholder that contains offset and size to the binary output.

Note the server could take the serialized version of JSON-bag and serve it.

HTTP 2.0

In HTTP 2.0 there is the possibility of Server Push that is sending additional content together with a response. In this scenario JSON-bag is straightforward: the server produces the first answer as JSON containing the references to the (virtual) URL pushed together the JSON.


Normative References

And notes:

Binary Manipulation in JS

Building Note

Example is with mongoose but it is not strictly necessary


  • size limit for data: URI (32k in some browsers)
  • URL split mode in server (client requires no changes)
  • emit chunked HTTP (JSON and then each blob)
  • progressive
  • deserialization in C++

See also the Issue Tracker


BInary multipart JSON for efficient payload




No releases published


No packages published