DICE API Documentation/com.embabel.dice.projection.memory.collector

Package-level declarations

Types

data class CollectorRunContext(val runId: String, val contextId: <Error class: unknown class>, val dryRun: Boolean = false, val clock: Clock = Clock.systemUTC())

Everything one run of the multi-signal collector needs beyond the candidate propositions themselves: a run id to tag its trace rows with, which context it's sweeping, whether it should hold off on anything that isn't read-only, and a clock so tests can pin "now".

CollectorSurvivorPolicy

fun interface CollectorSurvivorPolicy

Picks the one proposition in a component that survives a merge; everyone else in the component gets folded into it. Pulled out as a seam so a caller can plug in a different policy.

EntityOverlapSignalScorer

class EntityOverlapSignalScorer constructor(weight: Double = 1.0) : CollectorSignalScorer

Scores how much two propositions' entity mentions overlap, as Jaccard similarity of their stable mention ids.

GroundingOverlapSignalScorer

class GroundingOverlapSignalScorer constructor(weight: Double = 0.5) : CollectorSignalScorer

Scores how much two propositions' grounding chunk ids overlap, as Jaccard similarity. Two propositions extracted from the same source chunks are more likely restating the same fact.

LexicalSignalScorer

class LexicalSignalScorer constructor(weight: Double = 0.5) : CollectorSignalScorer

Scores how similar two propositions' surface text is, using Commons Text's Jaro-Winkler similarity.

MultiSignalCollectorStrategy

class MultiSignalCollectorStrategy(pairSources: List<CandidatePairSource>, scorers: List<CollectorSignalScorer>, componentsFinder: ConnectedComponentsFinder, traceStore: CollectorTraceStore, survivorPolicy: CollectorSurvivorPolicy = defaultCollectorSurvivorPolicy, matchThreshold: Double = 0.6) : RunAwareCollectorStrategy

Composes any number of candidate-pair sources and signal scorers into one duplicate-collapse strategy, replacing the single-cosine cut in com.embabel.dice.projection.memory.DuplicateCollectorStrategy with a pluggable multi-signal blend.

PolarityVetoSignalScorer

class PolarityVetoSignalScorer : CollectorSignalScorer

Vetoes a merge when two propositions share an entity but assert opposite polarity about it — one states something, the other negates it ("works at Acme" vs "no longer works at Acme" / "left Acme").

ProvenanceOverlapSignalScorer

class ProvenanceOverlapSignalScorer constructor(weight: Double = 0.5) : CollectorSignalScorer

Scores how much two propositions' rich provenance overlaps, as Jaccard similarity of their com.embabel.dice.provenance.SourceLocators. Locators compare by structural equality, so this catches "same source material" even when grounding chunk ids differ or are absent.

VectorCandidatePairSource

class VectorCandidatePairSource(repository: PropositionRepository, similarityThreshold: Double = 0.7, topK: Int = 10) : CandidatePairSource

Proposes candidate pairs from the repository's own vector clustering, the same recall path com.embabel.dice.projection.memory.DuplicateCollectorStrategy uses today.

VectorSignalScorer

class VectorSignalScorer constructor(weight: Double = 1.0) : CollectorSignalScorer

Turns the cosine similarity a VectorCandidatePairSource already computed into a signal score, without re-embedding anything.

Properties

defaultCollectorSurvivorPolicy

val defaultCollectorSurvivorPolicy: CollectorSurvivorPolicy

Default CollectorSurvivorPolicy: highest effective confidence, then reinforcement count, then stable id, so a full tie is still deterministic.