Explainability

Import the API collection into Bruno, Postman, or Insomnia to try all examples locally.

SAP-RPT-1.6 supports an explainability feature that provides insight into how the model arrives at its predictions. Enable it by adding an explanations object to the request body.

Enabling Explanations

Add the explanations field to your request:

{
  "rows": [...],
  "explanations": {
    "top_column_scores": 4,
    "top_relevant_context_rows": 3
  }
}

Parameters:

Parameter	Type	Range	Default	Description
`top_column_scores`	integer	1–20	4	Number of top feature importance scores to return per query row
`top_relevant_context_rows`	integer	1–20	3	Number of most relevant context row indices to return per query row

Response Format

When enabled, the explanations field in the response contains:

{
  "explanations": {
    "top_column_scores": [
      {
        "CATEGORY": 0.61,
        "PRODUCT": 0.28,
        "REGION": 0.11
      }
    ],
    "top_relevant_context_rows": [
      [4, 1, 7]
    ]
  }
}

Response Fields

Field	Type	Description
`explanations`	object \| null	Present only when explainability is enabled. `null` when not requested.
`explanations.top_column_scores`	array \| null	Array of objects (one per query row), mapping feature column names to attention-based importance scores. Higher score = more weight. Only top-k columns returned.
`explanations.top_relevant_context_rows`	array \| null	2D array — each sub-array lists zero-indexed indices of the most relevant context rows for that query row, sorted by relevance (highest first).

Notes

Column scores and context row indices are derived from the model's internal attention representations and can behave unintuitively when:

The dataset is very small
Feature columns are highly redundant or correlated

When multiple features are strongly correlated, the model may distribute scores across them rather than assigning a high score to a single column.