justmytyp#

An HTTP server for Typst that compiles preconfigured templates with input from HTTP requests. It replies with a PDF or uploads it to S3.

The input field in the request JSON is injected as a virtual file called input.json in the compilation context. In the template it can be accessed with json("/input.json").¹

Run in this repo, and try the following request:

(http post localhost:3000/
  --content-type application/json
  --full
  {
    name: "foo",
    template: "src/tests/test.typ",
    input: {
      title: "foo",
      text: "bar"
    }
  }
) | get body | save tmp.pdf

The compilation context is reused, so the speed should be comparable with typst watch (I think). Hyperfine says 2.3 ms ± 2.8 ms (meaning it's faster than the margin for error in its measurement of the shell startup time), and Apache's server benchmarking tool completes 1000 requests in under a second (using ab -n 1000 -c 1000).

Configuration is done by environment variables. See the Config struct and its Default implementation for the variables and their defaults.

To-do#

Soon#

In addition to the in-text notes, the following are priorities.

• Avoid unnecessary allocations.
• Add/fix tests.
• Investigate startup time lag.
• NixOS module.

Later#

These are ideas for improvements.

• Allow customisation of S3 folder structure (e.g. per date).
• Support document formats other than PDF.
• Refresh fonts (font library is currently only built once at startup).
• Add tracing logging?
• Endpoint to list available templates?
• Handle large files in response and multipart s3 upload.
• OT/Prometheus endpoint?
• Authorisation?
• Rate-limiting?

Credit#

The implementation of the compiler context and the interaction with the Typst libraries was based on tfachmann's example code.

¹I considered using CLI inputs (--inputs / sys.inputs), but it is surprisingly tedious to convert the JSON string to the expected Typst-internal types. With the virtual file, template development is also simpler --- especially when dealing with large inputs.