Member Junction
    Preparing search index...

    Options passed per Classify() call.

    interface ClassifyOptions {
        confidenceFloor?: number;
        fields: MJIntegrationObjectFieldEntity[];
        llmInference?: LLMOneShotCallback;
        maxCompositeKeySize?: number;
        object: MJIntegrationObjectEntity;
        sampleRows?: Record<string, unknown>[];
        syntheticFallback?: boolean;
        universalConvention?: string;
    }
    Index

    Properties

    confidenceFloor?: number

    Confidence floor for Confident=true. Defaults to 0.7.

    All IOF rows for that object.

    llmInference?: LLMOneShotCallback

    Optional LLM callback for the one-shot fallback. When omitted, the LLM step is skipped and the classifier returns Confident=false rather than guessing.

    maxCompositeKeySize?: number

    Largest column-set size the composite-key uniqueness scan will try when no single column is unique. Defaults to 3 (try 2-col then 3-col sets). Only consulted when sampleRows are present.

    The IntegrationObject being classified.

    sampleRows?: Record<string, unknown>[]

    Optional pre-fetched sample rows for statistical uniqueness check. When omitted, the statistical step is skipped (no chicken-and-egg fetching).

    syntheticFallback?: boolean

    When true (default), and every other tier fails to find/prove a PK, the classifier returns a Confident synthetic verdict nominating SYNTHETIC_PK_FIELD_NAME — an identity-hash column the framework materializes so NO table is left without a PK (plan.md §4). Set false to opt out and receive the honest 'none' verdict instead.

    universalConvention?: string

    Optional vendor-wide universal PK hint (e.g. "all HubSpot object PKs are id"). When set and a matching field exists, returned with strategy='universal-convention' and high confidence. Agent populates this from research.