Skip to main content
Version: latest

agrag.ingestion.settings.ExtractionSettings

Bases: BaseSettings

Configuration for local entity and relation extraction.

Every score is between 0 and 1. ExtractionLLMSettings variables start with EXTRACTION_LLM_ and are not read here, so no field may start with llm_.

The entity and relation thresholds, the relation floor, the entity band and the candidate switch shape what GlinerExtractor keeps and reports. The escalation threshold, budget and minimum word count decide which weak chunks go to the LLM extractor in hybrid mode.

Attributes:

  • mode (Literal['gliner', 'hybrid', 'llm']) – Extractor mode. gliner runs GLiNER only; hybrid runs GLiNER and may escalate weak chunks to the LLM; llm runs the LLM on every chunk. Escalation settings apply only to hybrid. This mode does not control entity resolution. Resolution has its own optional LLM tier, which runs whenever LLM settings exist, even in gliner mode.
  • model_name (str) – GLiNER checkpoint.
  • dtype (Literal['fp32', 'fp16', 'bf16']) – Model precision: fp32, fp16 or bf16.
  • entity_threshold (_Score) – Minimum score for a kept entity span.
  • relation_threshold (_Score) – Minimum score for a kept relation.
  • relation_floor (_Score) – Minimum relation score returned by the model.
  • band_floor (_Score) – Minimum entity score in the band below the keep threshold.
  • label_thresholds (dict[str, _Score]) – Per-label thresholds keyed by entity or relation label. An entity label's threshold replaces entity_threshold for that label. A relation label's threshold replaces relation_threshold as the cut for kept relations, and the model reads that relation at the lower of the threshold and relation_floor.
  • escalation_threshold (_Score) – Entity score below which a chunk may use the LLM.
  • escalation_budget (_Score | Literal['all']) – LLM chunk share from 0 to 1, or "all" for no limit.
  • min_escalation_words (int) – Minimum words for a chunk with no entities to use the LLM.
  • candidates_enabled (bool) – Whether discarded items are kept as candidates.

Env prefix: EXTRACTION_. label_thresholds uses one JSON object, such as EXTRACTION_LABEL_THRESHOLDS='{"Person": 0.7}'.

band_floor​

band_floor: _Score = 0.3

candidates_enabled​

candidates_enabled: bool = True

dtype​

dtype: Literal['fp32', 'fp16', 'bf16'] = 'fp32'

entity_threshold​

entity_threshold: _Score = 0.5

escalation_budget​

escalation_budget: _Score | Literal['all'] = 0.2

escalation_threshold​

escalation_threshold: _Score = 0.75

label_thresholds​

label_thresholds: dict[str, _Score] = Field(default_factory=dict)

min_escalation_words​

min_escalation_words: int = Field(default=8, ge=0)

mode​

mode: Literal['gliner', 'hybrid', 'llm'] = 'hybrid'

model_config​

model_config = SettingsConfigDict(env_prefix='EXTRACTION_', env_file='.env', extra='ignore', hide_input_in_errors=True, protected_namespaces=())

model_name​

model_name: str = 'fastino/gliner2.5-base-v1'

relation_floor​

relation_floor: _Score = 0.3

relation_threshold​

relation_threshold: _Score = 0.7