Proposition
Pipeline for extracting propositions from chunks. Coordinates extraction and entity resolution.
Construction
There is no public constructor. The only entry point is the companion factory withExtractor, which seeds a pipeline with a PropositionExtractor. From there, configure the pipeline with the fluent copy-builders, each of which returns a new instance:
withRevision — compare new propositions against existing ones
withMentionFilter — drop low-quality entity mentions before resolution
withEventListener — observe pipeline events
withExecutionStrategy — control how the extraction stage is dispatched
Example usage:
val pipeline = PropositionPipeline
.withExtractor(LlmPropositionExtractor(ai))
.withRevision(reviser, propositionRepository) // Optional
val result = pipeline.process(chunks, context)This pipeline does NOT persist anything
process, processChunk, and processOnce all return UNSAVED results. The pipeline writes nothing to any repository on its own. To make results durable, the caller MUST persist them explicitly via PersistablePropositions.persist, passing a propositionRepository and a namedEntityDataRepository, within the caller's own transaction scope. If you do not call persist, nothing is stored — the returned result is discarded when it goes out of scope.
When a PropositionReviser is configured with a PropositionRepository, the pipeline will compare new propositions against existing ones and classify them as new, merged, reinforced, or contradicted. Revision still does not persist — the classified results must be persisted by the caller as above.
Functions
Process multiple chunks through the pipeline.
Process a single chunk through the pipeline. Extracts propositions and resolves entities.
Process a text once, with hash-based deduplication. Ideal for one-shot ingestion of documents, notes, or other static text.
Register a listener for pipeline events.
Set the ExtractionExecutionStrategy used to dispatch the stateless per-chunk extraction stage of process.
Add a mention filter to validate entity mentions before creating entities. When enabled, low-quality mentions (vague references, overly long spans, etc.) are filtered out before entity resolution.
Add a reviser to compare new propositions against existing ones. When enabled, propositions are classified as new, merged, reinforced, or contradicted.