Vectorize - Verglas

Vectorize is a named Durable Object with vector CRUD, namespace and metadata filtering, and similarity search. Create an index with fixed dimensions and one metric: cosine, euclidean, or dot-product.

Bind an index

{
  "vectorize": [
    { "binding": "DOCUMENTS", "index_name": "product-docs" }
  ]
}

Insert or update vectors

await env.DOCUMENTS.upsert([
  {
    id: "doc-42",
    values: embedding,
    namespace: "guides",
    metadata: { title: "Deploy a Worker", locale: "en" },
  },
]);

insert preserves an existing ID; upsert fully replaces it. The mutation is visible when the promise resolves.

Query

const result = await env.DOCUMENTS.query(queryEmbedding, {
  topK: 5,
  namespace: "guides",
  returnMetadata: "all",
  filter: { locale: "en" },
});

You can also search using a stored vector:

const similar = await env.DOCUMENTS.queryById("doc-42", { topK: 5 });

Complete API

Method	Purpose
`insert(vectors)`	Add IDs without replacement
`upsert(vectors)`	Add or replace IDs
`query(vector, options)`	Find nearest vectors
`queryById(id, options)`	Use an existing vector as the query
`getByIds(ids)`	Fetch complete vectors
`deleteByIds(ids)`	Remove vectors
`describe()`	Inspect immutable index configuration and progress

Requests accept ordinary arrays, Float32Array, or Float64Array. Query topK is between 1 and 100; batch ID operations accept at most 1,000 IDs.

Vectorize performs exact native Turso similarity search. A query with more than 10,000 eligible rows fails explicitly instead of silently switching to an approximate or remote index.