DocumentClassificationSettings
Settings for DocumentClassification. Values fall back through three levels: document → SDK → built-in default. Writes target the document only when set on a document’s settings, otherwise the SDK globally when set on SdkSettings.
using Nutrient;The SDK creates this class through factory methods or other SDK objects.
Properties
AutoWeight
public bool AutoWeight { get; set; }Auto (reliability) weighting. When true (default), the text/image split is derived PER DOCUMENT from each branch’s decisiveness (the top1−top2 margin of its per-label scores) — the more confident branch on that document is trusted more — and TextWeight/ ImageWeight are ignored. When false, the fixed TextWeight/ ImageWeight ratio is used. In BOTH modes a weight of 0 still fully disables that branch (skipping its extraction/rendering). The effective per-document image weight is reported on the result for transparency.
Type: bool
EnableDocumentDecisionClassifier
public bool EnableDocumentDecisionClassifier { get; set; }When true, classify via the nutrient-document-decision GGUF model instead of the SigLIP2/bge-m3/learned-fusion specialist — a full substitute backend, not a third branch fused with the other two (TextWeight/ImageWeight/ AutoWeight are ignored; MultiLabel still selects the question shape). Resolved from the CDN. Slower than the specialist — opt-in, not a default. Default false.
Type: bool
ImageWeight
public float ImageWeight { get; set; }Relative weight of the image branch in the fused prediction. Set to 0 to use the text branch alone. Default 0.5. Only consulted when AutoWeight is false.
Type: float
IncludePerPageScores
public bool IncludePerPageScores { get; set; }Diagnostic opt-in. When true, the classification artifact additionally carries, per ranked prediction, the raw per-page image-branch confidences before multi-page pooling (imageProbabilityPerPage) — letting a caller reconstruct any pooling strategy or PoolingTemperature offline without re-running the model. Default false so production artifacts stay lean; the field is omitted entirely when off.
Type: bool
MultiLabel
public bool MultiLabel { get; set; }Confidence mode (reporting only — does not change the predicted label or ranking). True (default): INDEPENDENT per-label confidence, each in [0,1] (like HuggingFace multi_label=True) — supports multi-label (threshold each label) and out-of-category abstention. False (single-label): the reported per-label confidences are normalized to sum to 1 (share-of-1). The predicted label, ranking, and returned label set are identical either way; only the confidence scale changes. Abstention (the fusion action’s confidenceThreshold input, not a setting) always uses the raw independent confidence.
Type: bool
PoolingTemperature
public float PoolingTemperature { get; set; }Margin-weighted pooling temperature used by both branches to combine the per-chunk (text) and per-page (image) distributions into one. Higher values let the most distinctive chunk/page dominate more aggressively; values approaching 0 approach a plain average. Default 8.
Type: float
Text
public string Text { get; set; }Document text for the text branch. When non-empty, this text is scored directly. When empty (default), the text is derived from the document with AdaptiveOcr (fast OCR), or with Engine when it is set (e.g. Icr for local layout models, VlmEnhancedIcr for VLM-enhanced) — multi-page documents are extracted in full. Set TextWeight to 0 to skip the text branch (and its extraction) entirely.
Type: string
TextWeight
public float TextWeight { get; set; }Relative weight of the text branch in the fused prediction. The fused probability for a label is (TextWeight·Pt + ImageWeight·Pi) / (TextWeight + ImageWeight). Set to 0 to use the image branch alone. Default 0.5.
Type: float
TopK
public int TopK { get; set; }Maximum number of ranked predictions to return. 0 (default) returns all candidates, ranked by fused probability descending.
Type: int
Resource management
public void Dispose()DocumentClassificationSettings implements IDisposable. Call Dispose() or use a C# using declaration to release its native handle right away. If you don’t, the handle is released when the object is garbage collected.