Extractor builder
Build a local extractor from enabled stages, from Extractor classes reference. Import ExtractorBuilder from neo4j_agent_memory.extraction.
ExtractorBuilder
Build a local extractor from enabled stages. build() returns NoOpExtractor when no stage is enabled, the concrete extractor for one stage, or ExtractionPipeline for multiple stages.
from neo4j_agent_memory.extraction import ExtractorBuilder
extractor = (
ExtractorBuilder()
.with_spacy("en_core_web_sm")
.with_gliner_schema("podcast", threshold=0.5)
.with_llm_fallback("gpt-4o-mini")
.merge_by_confidence()
.build()
)The union method is merge_by_union(). with_gliner_relations() was removed in 0.7: the GLiNER2.5 stage added by with_gliner or with_gliner_schema decodes relations itself.
build() resolves one ontology and gives it to every stage it creates: the spaCy label mapping, the GLiNER2.5 JointIE schema, the LLM prompt, and the pipeline’s relation validation. A with_gliner_schema template takes precedence over a document from with_ontology or with_schema. When no ontology is named, the GLiNER2.5 stage uses POLE+O (or the with_entity_types labels) and a multi-stage build does not validate relations.
with_spacy
Add spaCy extractor to pipeline.
def with_spacy(model: str='en_core_web_sm') -> ExtractorBuilder: ...with_gliner
Add the GLiNER2.5 stage on the builder’s ontology (POLE+O when none is set).
def with_gliner(
model: str | None=None,
threshold: float=0.5,
device: str='cpu',
*,
model_name: str | None=None,
relation_threshold: float | None=None,
) -> ExtractorBuilder: ...| Parameter | Description |
|---|---|
|
GLiNER2.5 checkpoint (default |
|
Entity confidence threshold. |
|
Inference device ( |
|
Alias for |
|
Confidence floor for decoded relations. |
"gliner" in the builder method names means GLiNER2.5.
with_gliner_schema
Add the GLiNER2.5 stage on a built-in domain template. The template carries entity descriptions, and poleo, podcast and news also carry typed relationships.
def with_gliner_schema(
schema_name: str,
model: str | None=None,
threshold: float=0.5,
device: str='cpu',
*,
relation_threshold: float | None=None,
) -> ExtractorBuilder: ...| Parameter | Description |
|---|---|
|
Name of the template (poleo, podcast, news, scientific, business, entertainment, medical, legal), resolved with |
|
GLiNER2.5 checkpoint; |
|
Entity confidence threshold |
|
Device to run on (cpu, cuda, mps) |
|
Confidence floor for decoded relations |
with_ontology
Extract against an explicit OntologyDocument. This configures the ontology only; add the stages with with_gliner, with_llm_fallback or with_spacy. The LLM stage’s flat type set becomes the ontology’s POLE+O types.
def with_ontology(ontology: OntologyDocument) -> ExtractorBuilder: ...from neo4j_agent_memory.extraction import ExtractorBuilder
from neo4j_agent_memory.ontology import get_template
extractor = ExtractorBuilder().with_ontology(get_template("news")).with_gliner().build()with_llm
Add an LLM extractor (alias of with_llm_fallback).
def with_llm(model: str='gpt-4o-mini') -> ExtractorBuilder: ...with_llm_fallback
Add an LLM extractor stage. Despite the name, the stage is not a fallback: the pipeline that build() returns runs every stage on every call and merges the results with the selected merge strategy, so the LLM is called for every text even when spaCy or GLiNER2.5 find entities. The builder has no early-stop option. For true fallback behaviour, construct ExtractionPipeline directly with stop_on_success=True or MergeStrategy.FIRST_SUCCESS.
def with_llm_fallback(model: str='gpt-4o-mini') -> ExtractorBuilder: ...with_entity_types
Set the flat entity type set for the LLM stage. When neither an ontology nor a template is set, the lower-cased types also become the GLiNER2.5 labels; that ad-hoc label set declares no relationships, so the GLiNER2.5 stage then decodes entities only.
def with_entity_types(types: list[str]) -> ExtractorBuilder: ...with_schema
Configure entity types from an EntitySchemaConfig.
def with_schema(schema_config: EntitySchemaConfig) -> ExtractorBuilder: ...| Parameter | Description |
|---|---|
|
Entity schema configuration. Each |
The schema is converted with EntitySchemaConfig.to_ontology(), which becomes the builder’s ontology: entity types and, when enable_subtypes is on, their subtypes become GLiNER2.5 labels, and relation_types become typed relationships expanded over source_types × target_types. Endpoints naming a type the schema does not declare are dropped. It does not create constraints or indexes, and it does not store or activate anything.
with_confidence_threshold
Set the confidence threshold for entity extraction. GLiNER2.5 decodes with it as its entity threshold, and when build() returns a pipeline, entities and relations below it are dropped from the merged result, so spaCy and LLM stages are held to the same floor. A value outside [0.0, 1.0] raises ValueError.
def with_confidence_threshold(threshold: float) -> ExtractorBuilder: ...merge_by_intersection
Use intersection strategy for merging.
def merge_by_intersection() -> ExtractorBuilder: ...merge_by_confidence
Use confidence strategy for merging.
def merge_by_confidence() -> ExtractorBuilder: ...extract_relations
Enable/disable relation extraction.
def extract_relations(enabled: bool=True) -> ExtractorBuilder: ...