Vector stores

A vector store holds embedded documents and exposes a provider-neutral VectorSearchIndex to the agent runtime.

Build a local index

import { embedDocuments } from '@anvia/core/embeddings'
import { InMemoryVectorStore } from '@anvia/core/vector-store'

const embedded = await embedDocuments(embeddingModel, documents, {
  id: (document) => document.id,
  content: (document) => document.text,
  metadata: (document) => ({
    source: document.source,
    product: document.product,
  }),
})

const localStore = InMemoryVectorStore.fromDocuments(embedded)
const index = localStore.index(embeddingModel)

The in-memory store is suitable for tests, demos, and small process-local indexes.

Search the index

const results = await index.search({
  query: 'How long does a password reset link last?',
  topK: 3,
  threshold: 0.72,
})

Results are ordered by descending score and include the document ID, original document, and optional metadata. A threshold removes weak matches before they reach the model.

Use searchIds(...) when the index should identify likely records and the application must load their full contents through its own data-access layer.

Update an in-memory store

const localStore = InMemoryVectorStore.fromDocuments(embedded)

const updatedDocuments = await embedDocuments(
  embeddingModel,
  changedDocuments,
  {
    id: (document) => document.id,
    content: (document) => document.text,
  },
)

localStore.addDocuments(updatedDocuments)

The in-memory store replaces an existing document with the same ID. Stable IDs make updates predictable and prevent old chunks from remaining searchable.

Upsert production documents

Production adapters use the asynchronous, plural upsertDocuments(...) method:

import { QdrantVectorStore } from '@anvia/qdrant'

const documentsToUpsert = await embedDocuments(
  embeddingModel,
  changedDocuments,
  {
    id: (document) => document.id,
    content: (document) => document.text,
  },
)

const vectorStore = await QdrantVectorStore.connect({
  collectionName: 'support_docs',
  vectorSize: embeddingModel.dimensions,
})

await vectorStore.upsertDocuments(documentsToUpsert)

There is no singular upsertDocument(...) API. Pass one or several embedded documents as an array.

Choose a production adapter

Anvia provides adapters for pgvector, Qdrant, Pinecone, Chroma, LanceDB, Milvus, Redis, and Weaviate.

Keep credentials, collection names, and ingestion jobs outside the agent. Pass a prepared VectorSearchIndex into the agent factory.