Skip to main content
Version: latest

Turn candidates off

By default, Graph.add and Graph.update write discarded extraction items as candidates. The extraction and merge statistics still count discarded items when candidate writes are off. For candidate details, see Candidates.

Disable candidate writes​

Set the environment variable before you open the graph:

export EXTRACTION_CANDIDATES_ENABLED=false

Or pass one ExtractionSettings object to the graph and the GLiNER extractor. Install the neo4j and extract extras and start Neo4j first.

import asyncio

from agrag.common.data_models import DEFAULT_SCHEMA
from agrag.embedding import build_embedder
from agrag.graphdb import build_graph_store
from agrag.ingestion import ExtractionSettings, Graph, GlinerExtractor


async def main() -> None:
    graph_store = build_graph_store("neo4j")
    settings = ExtractionSettings(candidates_enabled=False)
    graph = await Graph.open(
        schema=DEFAULT_SCHEMA,
        graph_store=graph_store,
        embedder=build_embedder(
            "ibm-granite/granite-embedding-small-english-r2"
        ),
        extractor=GlinerExtractor(settings=settings),
        extraction_settings=settings,
    )
    try:
        result = await graph.add(text="A sample document with an uncertain relation.")
        print(result.extraction.candidates_written)
        for item in result.extraction.dropped:
            print(item.kind, item.label, item.reason, item.count)
    finally:
        await graph_store.close()


asyncio.run(main())

The script prints 0 for candidates_written. It also prints the discarded-item counts. The entity_band reason is absent because GLiNER does not run its candidate pass.

Existing candidates​

Turning candidate writes off does not delete existing candidates. Graph.update replaces a document's pending candidates with those from its latest run. Graph.delete_document removes that document's pending candidates. Approved and rejected candidates stay.