FormRecognitionSettings

Settings for FormRecognition. 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.

Tags: Vision, Advanced

from nutrient_sdk import FormRecognitionSettings

Construction

# Per-document override
with Document.open("input.pdf") as doc:
    settings = doc.settings.form_recognition_settings
    settings.some_field = new_value          # mutate fields directly

# SDK-wide default (applies to all documents)
SdkSettings.form_recognition_settings.some_field = new_value

Settings are configured by writing to fields on the returned object. The settings property itself cannot be reassigned — doc.settings.form_recognition_settings = other_settings is rejected.

Properties

checkbox_logit_bias

@property
def checkbox_logit_bias(self) -> float

@checkbox_logit_bias.setter
def checkbox_logit_bias(self, value: float) -> None

Logit bias applied to the Checkbox class before sigmoid during postprocessing. Positive values increase Checkbox detections, negative values suppress them. Default 0 (no bias).

Type: float

Default: 0f

confidence_threshold

@property
def confidence_threshold(self) -> float

@confidence_threshold.setter
def confidence_threshold(self, value: float) -> None

Confidence threshold for form field predictions that always exist as real fields, independent of the Specialist form-field VLM. This value has moved twice before (0.5 -> 0.35 -> 0.2), each time lowered because the 3-class form-field-v1 medium detector was observed to surface real-form fields at lower confidence than the synthetic benchmark implied. Default 0.35 (not the original, never-validated-on-real-documents 0.5 spec default) restores some of that precision, which is only safe because low_confidence_threshold now exists to recover genuinely weak-but-real fields when the VLM independently corroborates one geometrically (see FormVlmCorrelator’s three-phase matching) — raising this value without that corroboration path would reintroduce the exact recall loss this setting was lowered twice to avoid.

Type: float

Default: 0.35f

low_confidence_threshold

@property
def low_confidence_threshold(self) -> float

@low_confidence_threshold.setter
def low_confidence_threshold(self, value: float) -> None

A second, much lower confidence floor for form field predictions that must NEVER be trusted on their own and must NEVER be written to the DocumentGraph — candidates between this value and confidence_threshold only ever become real fields when the Specialist form-field VLM independently corroborates one geometrically (FormVlmCorrelator phase 3). This is what lets confidence_threshold sit higher (fewer uncorroborated false positives) without losing recall on genuinely real fields the detector alone is unsure about. Must be <= confidence_threshold; default 0.1 (tuned after field testing).

Type: float

Default: 0.1f

model_path

@property
def model_path(self) -> str

@model_path.setter
def model_path(self, value: str) -> None

Path to the ONNX RF-DETR model file (model.onnx).

Type: str

Default: GetDefaultModelPath()

num_select

@property
def num_select(self) -> int

@num_select.setter
def num_select(self, value: int) -> None

Number of top predictions to select during postprocessing. Default is 300 as specified in the RF-DETR configuration.

Type: int

Default: 300

signature_logit_bias

@property
def signature_logit_bias(self) -> float

@signature_logit_bias.setter
def signature_logit_bias(self, value: float) -> None

Logit bias applied to the Signature class before sigmoid during postprocessing. Positive values increase Signature detections, negative values suppress them. Default 0 (no bias).

Type: float

Default: 0f

text_logit_bias

@property
def text_logit_bias(self) -> float

@text_logit_bias.setter
def text_logit_bias(self, value: float) -> None

Logit bias applied to the Text class before sigmoid during postprocessing. Positive values increase Text detections, negative values suppress them. Default 0 (no bias).

Type: float

Default: 0f

use_cpu_only

@property
def use_cpu_only(self) -> bool

@use_cpu_only.setter
def use_cpu_only(self, value: bool) -> None

Enable CPU mode for ONNX. When true, forces CPU execution. When false, attempts to use GPU if available (CUDA/MPS).

Type: bool

Default: false