Minify introspect return value

[pubmodmatt]

Add options to minify the type information returned by the introspect and search tools. This emulates the GSL format defined here in Rust. This removes extraneous whitespace and braces, and shortens built-in type names and keywords. It also currently removes directives, though these may be considered useful to the LLM in some cases.

To enable minification, turn on the minify option in the config file, for example:

introspection:
  introspect:
    enabled: true
    minify: true
  search:
    enabled: true
    minify: true

For the top level Query type in the weather example schema, the original SDL was:

type Query {
  """Get the weather forecast for a coordinate"""
  forecast(coordinate: InputCoordinate!): Forecast
  """Get the weather alerts for a state"""
  alerts(
    """The two-letter state abbreviation (e.g., 'CO' for Colorado)"""
    state: String!,
  ): [Alert]
}

Minified, this becomes:

T:Query:"Gettheweatherforecastforacoordinate"forecast(InputCoordinate!):Forecast,"Gettheweatheralertsforastate"alerts("Thetwo-letterstateabbreviation(e.g.,'CO'forColorado)"state:s!):[Alert]

This reduces the size from 272 -> 190 bytes, about a 30% reduction.

For a more real-world example, the Repository type in the GitHub schema went from 35322 -> 24418 bytes, about a 31% reduction.

[apollo-librarian]

✅ Docs preview ready

The preview is ready to be viewed. View the preview

File Changes

0 new, 1 changed, 0 removed

* (developer-tools)/apollo-mcp-server/(latest)/quickstart.mdx

Build ID: 872015a8b89f2bc681480b37

URL: https://www.apollographql.com/docs/deploy-preview/872015a8b89f2bc681480b37

[alocay]

Do you have any estimate on how much this reduces the instrospect return value by?

[pubmodmatt]

Do you have any estimate on how much this reduces the instrospect return value by?

I've updated the PR description with some examples.

[DaleSeo]

I can't understand the complex logic here... 😵

[pubmodmatt]

Unfortunately apollo-compiler ergonomics generally result in a lot of code duplcation since the various types don't have traits in common. But this is really pretty simple - instead of using the ExtendedType.serialize() method, which outputs the full GraphQL SDL, this alternative ExtendedType.minify() method returns the smaller form for each type. So for each variant, it's building up the small string, which includes descriptions, field names, and arguments.

[pubmodmatt]

@DaleSeo - I refactored the minify.rs file to (hopefully) make it simpler to read and maintain.

[swcollard]

Is this how the LLM knows how to un-minify to execute actual queries? Or does it not need that information because the type and field names are unchanged?

[pubmodmatt]

Yes, exactly! This tiny bit of text is all the model needs to undo the minification. A big caveat here is that I haven't done extensive testing of this across different models, and only really tested with Claude Sonnet 4. Smaller models may struggle with it. I consider this feature to be experimental for that reason. There are other ways more detailed decoding information could be given to a model, such as through resources or system prompts. Or if the spec for the minified format were publicly documented, eventually models would have it as part of their training.