Create a Vectorize index sized for your embeddings, then insert/upsert and query vectors with topK limits and metadata filtering from a Worker.

Documented steps

1. Create an index with a fixed dimension size and distance metric: npx wrangler vectorize create <index-name> --dimensions=<N> --metric=<cosine|euclidean|dot-product> (e.g. --dimensions=768 --metric=cosine for Workers AI's @cf/baai/bge-base-en-v1.5). Both are immutable after creation. https://developers.cloudflare.com/vectorize/best-practices/create-indexes/
2. Bind the index to your Worker: {"vectorize": [{"binding": "VECTORIZE", "index_name": "<index-name>"}]} in the wrangler config, available in code as env.VECTORIZE. https://developers.cloudflare.com/vectorize/get-started/intro/
3. Before inserting data, create a metadata index for any field you plan to filter on: npx wrangler vectorize create-metadata-index <index-name> --property-name=<field> --type=<string|number|boolean> (max 10 per index). https://developers.cloudflare.com/vectorize/get-started/intro/
4. Insert vectors with await env.VECTORIZE.insert([{ id, values, metadata, namespace? }, ...]); on an id collision insert() keeps the first-written vector, so use upsert() instead if you want the newest value to win. https://developers.cloudflare.com/vectorize/best-practices/insert-vectors/
5. Query with const matches = await env.VECTORIZE.query(queryVector, { topK, returnValues, returnMetadata, namespace, filter }); topK defaults to 5, returnValues defaults to false, returnMetadata defaults to "none". https://developers.cloudflare.com/vectorize/best-practices/query-vectors/
6. Apply metadata filters via the filter option, e.g. { someKey: { "$eq": "value" } }, supporting $eq/$ne/$in/$nin/$lt/$lte/$gt/$gte; only fields with a metadata index can be filtered. https://developers.cloudflare.com/vectorize/reference/metadata-filtering/
7. Use queryById("vector-id") to search for vectors similar to one already stored in the index, combining getById and query in one call. https://developers.cloudflare.com/vectorize/best-practices/query-vectors/
8. Check async write progress with npx wrangler vectorize info <index-name>, which reports processedUpToMutation, since inserts/upserts aren't guaranteed queryable immediately. https://developers.cloudflare.com/vectorize/get-started/intro/