@samesake/enrich

@samesake/enrich owns the enrichment and resolution brain. It is store-free and performs model I/O only through closures supplied by the consumer.

Public surface

• enrich and enrichRow execute typed stages, derive surfaces, and return explicit ready, quarantined, or failed row outcomes.
• scoreCandidate, scoreBest, and clusterBatch implement pure resolution scoring, threshold decisions, and offer grouping.
• contentHash and selectDirty support incremental work without owning persistence.
• scoreEnrichment and its result types score predictions against gold attributes.
• createEnricher composes those cores with an EnrichStore, optional candidate provider, generation closure, embedding closure, and consumer-chosen concurrency.
• memoryStore is a reference in-memory store for tests and local experiments.

Ports

EnrichStore is the durable row-state contract. It covers upsert, dirty loading, optional hard deletion, enriched writes, failures, retries, and optional enriched-row/candidate access for resolution. CandidateProvider supplies a shortlist; blocking is backend work, while scoring remains pure.

The factory returns an Enricher with:

upsert(rows)
remove(ids)
enrich({ limit, concurrency })
resolve({ limit })
retryFailed({ limit })
evaluate(gold)

remove(ids) calls the store’s optional hard-delete capability and throws a clear error when the store does not implement it. createEnricher does not schedule jobs, retry network calls, or run a dead-letter queue. The application platform and its store own those concerns.

Example

import { createEnricher } from '@samesake/enrich';

const enricher = createEnricher({
  collection,
  generate,
  embed,
  store,
});

await enricher.upsert(rows);
await enricher.enrich({ limit: 100 });

Use the pure functions directly when the surrounding workflow already owns durable state. Use the factory when one application-owned store should coordinate the loop.