Get started

ocd must be ready. Create a Vectorize index, then bind it from wrangler.jsonc.

1. Create an index

Use pinned Wrangler against live ocd (or the local Cloudflare v4 API):

wrangler vectorize create embeddings --dimensions=768 --metric=cosine

2. Declare the binding

{
  "name": "vector-app",
  "main": "src/index.ts",
  "vectorize": [{ "binding": "VECTORIZE", "index_name": "embeddings" }]
}

bun run oc types --config wrangler.jsonc

3. Worker

export default {
  async fetch(request: Request, env: Env): Promise<Response> {
    if (request.method === "PUT") {
      const body = await request.json<{ id: string; values: number[] }>();
      const { mutationId } = await env.VECTORIZE.upsert([
        { id: body.id, values: body.values, metadata: { source: "api" } },
      ]);
      return Response.json({ mutationId });
    }
    const { matches } = await env.VECTORIZE.query(
      await request.json<number[]>(),
      { topK: 10, returnMetadata: "all" },
    );
    return Response.json(matches);
  },
} satisfies ExportedHandler<Env>;

You supply vectors; open-compute does not generate embeddings. For document ingest + embedding + retrieval, see AI Search.

4. Deploy

bun run oc deploy --config wrangler.jsonc

Next: Concepts.