Embeddings

Use embeddingModel(...) to turn text into vectors for ingestion, retrieval, clustering, or semantic search.

import { OpenAIClient } from '@anvia/openai'

const openai = new OpenAIClient({
  apiKey: process.env.OPENAI_API_KEY,
})

export const embeddingModel = openai.embeddingModel(
  'text-embedding-3-small',
  {
    dimensions: 1536,
    maxBatchSize: 512,
    user: 'tenant-123',
  },
)

Embed application text

import { embedText, embedTexts } from '@anvia/core/embeddings'

const query = await embedText(embeddingModel, 'refund timeline')

const documents = await embedTexts(embeddingModel, [
  'Refunds are reviewed within two business days.',
  'Password reset links expire after 30 minutes.',
])

The adapter returns one embedding for each input, preserves input order, and rejects mismatched provider response counts or indexes rather than returning silently misaligned vectors.

Configure dimensions once

dimensions changes the vector width for models that support it. The value must match the vector-store schema used for both document ingestion and query embedding. Changing it normally requires a new index or a full re-embedding migration.

maxBatchSize controls how many texts the model receives per provider request. It is a request batching limit, not a concurrency limit. Bound worker concurrency separately to stay within provider and database capacity.

The optional user value is forwarded as provider metadata. Use a stable, non-sensitive identifier rather than an email address or raw customer name.

Build a retrieval boundary

Keep one configured embedding model with the index it belongs to:

export const supportIndex = vectorStore.index(embeddingModel)

Ingestion and query code should share that boundary so model ID, dimensions, normalization assumptions, and store schema cannot drift independently. See Knowledges for the complete ingestion and retrieval flow.