Get a JSON explanation

If your sample came from a JsonGrinder.jl extractor, you can reconstruct the pruned explanation as a JSON-shaped value – the original field values (element names, categories, numbers, strings) with pruned parts represented as nothing, instead of raw one-hot/matrix data.

Requirement: extract with store_input=Val(true)

This only works if the sample was extracted with metadata preserved:

ds = extractor(json_sample; store_input=Val(true))
# or, for a batch:
x = extract(extractor, samples; store_input=Val(true))

Without this, every leaf's .metadata is nothing, and explain_json raises a clear error the first time it needs a value it doesn't have:

explain_json: no metadata on a ArrayNode leaf -- re-extract with
`store_input=Val(true)` so explain_json has original values to reconstruct

If you forget until after training/scoring, just re-extract the one sample you want to explain – store_input=Val(true) is not needed for training data, only for whichever sample(s) you'll later call explain_json on.

Usage

using ExplainMillX, JSON

result = explain(ds, model)
explain_json(result, extractor)             # convenience: uses result.sample
explain_json(ds, result.mask, extractor)    # equivalent, spelled out

JSON.print(explain_json(result, extractor), 4)   # pretty-printed JSON string

extractor must be the same (or a structurally equivalent) extractor used to produce ds – reconstruction depends on knowing whether a field was a scalar, a category, an array, etc., which only the extractor knows.

What the output looks like

Pruned leaves are dropped (not shown as explicit null), and a field that's entirely pruned away disappears from its parent object rather than appearing as an empty value:

{
    "atoms": [
        {"charge": 0.812},
        {"element": "o", "charge": -0.388}
    ],
    "logp": 4.44
}

If everything was pruned, explain_json returns nothing rather than an empty object.

Limitations

• One sample at a time – ds must have numobs(ds) == 1.
• PolymorphExtractor (JsonGrinder's union-typed fields) is not supported and raises a clear error.