Collections and indexing

Automatic creation defines either an unnamed dense vector or named dense and sparse vectors. Production infrastructure should instead own collection creation, vector settings, replication, sharding, payload indexes, and snapshots.

const store = await QdrantVectorStore.connect({
  client,
  collectionName: 'support_docs_hybrid',
  vectorSize: 1536,
  hybrid: true,
  denseVectorName: 'dense',
  sparseVectorName: 'sparse',
  createIfMissing: false,
})

vectorSize, distance, and dense/hybrid shape must match the existing collection. connect(...) validates these settings whether automatic creation is enabled or disabled. Hybrid ingestion also validates that sparse embeddings align with dense embeddings. Metadata keys beginning with __anvia_ are reserved.

Replace and inspect logical documents

One logical document can produce several Qdrant points. upsertDocuments(...) removes all existing points for each incoming document ID before inserting its current embeddings, preventing stale points when chunking changes.

await store.upsertDocuments(documents, {
  wait: true,
  ordering: 'strong',
  timeout: 30,
})

await store.deleteDocuments(['obsolete-document'])

const firstPage = await store.index(embeddings).inspect({ limit: 50 })
const nextPage = firstPage.nextCursor
  ? await store.index(embeddings).inspect({
      limit: 50,
      cursor: firstPage.nextCursor,
    })
  : undefined

The official client performs replacement with batchUpdate(...). A narrow custom client may fall back to a sequential delete and upsert, which is not atomic if the upsert fails. Document retrieval and inspection require scroll(...) on custom clients.

Add payload indexes for fields used frequently in filters; the adapter does not create them.