Embeddings

Use embeddingModel(...) to create vectors for retrieval, semantic similarity, classification, clustering, and related workloads.

import { GeminiClient } from '@anvia/gemini'

const gemini = new GeminiClient({
  apiKey: process.env.GEMINI_API_KEY,
})

export const documentEmbeddings = gemini.embeddingModel(
  'gemini-embedding-001',
  {
    taskType: 'RETRIEVAL_DOCUMENT',
    dimensions: 768,
    maxBatchSize: 100,
  },
)

Embed text

const embeddings = await documentEmbeddings.embedTexts([
  'Refunds take five business days.',
  'Password reset links expire after 30 minutes.',
])

for (const embedding of embeddings) {
  console.log(embedding.document, embedding.vector.length)
}

The adapter preserves input order and returns one { document, vector } result for each input. It splits large input arrays into provider calls according to maxBatchSize, which defaults to 100.

Match task types

For retrieval, configure separate model objects for indexed documents and search queries:

const documentEmbeddings = gemini.embeddingModel(
  'gemini-embedding-001',
  {
    taskType: 'RETRIEVAL_DOCUMENT',
    dimensions: 768,
  },
)

const queryEmbeddings = gemini.embeddingModel(
  'gemini-embedding-001',
  {
    taskType: 'RETRIEVAL_QUERY',
    dimensions: 768,
  },
)

Keep the model ID and output dimensions identical between indexed documents and queries. Changing either requires re-embedding the stored collection.

Supported adapter task values are:

• RETRIEVAL_QUERY and RETRIEVAL_DOCUMENT
• SEMANTIC_SIMILARITY
• CLASSIFICATION and CLUSTERING
• QUESTION_ANSWERING and FACT_VERIFICATION
• CODE_RETRIEVAL_QUERY
• TASK_TYPE_UNSPECIFIED

The selected Google model must support the requested task type.

Model options

Option	Purpose
dimensions	Sends Google's output dimensionality and exposes it on the Anvia model.
maxBatchSize	Limits the number of texts in each provider request.
taskType	Tunes embeddings for the declared workload.
title	Supplies a document title in the embedding configuration.

Use title only where one title meaningfully applies to the texts sent through that model instance. For document collections with different titles, create appropriately scoped models or omit the shared option.

Production boundaries

Exclude secrets and private fields before embedding. Keep stable document IDs, permission metadata, model ID, dimensions, and task configuration with the ingestion version. Embeddings are derived data, but they still need access control and retention policies.