Cloudflare Vectorize

If you're deploying your project in a Cloudflare worker, you can use Cloudflare Vectorize with LangChain.js. It's a powerful and convenient option that's built directly into Cloudflare.

Setup

Compatibility

Cloudflare Vectorize is currently in open beta, and requires a Cloudflare account on a paid plan to use.

After setting up your project, create an index by running the following Wrangler command:

$ npx wrangler vectorize create <index_name> --preset @cf/baai/bge-small-en-v1.5

You can see a full list of options for the vectorize command in the official documentation.

You'll then need to update your wrangler.toml file to include an entry for [[vectorize]]:

[[vectorize]]
binding = "VECTORIZE_INDEX"
index_name = "<index_name>"

Finally, you'll need to install the LangChain Cloudflare integration package:

tip

See this section for general instructions on installing integration packages.

npm

npm install @langchain/cloudflare

Yarn

yarn add @langchain/cloudflare

pnpm

pnpm add @langchain/cloudflare

Usage

Below is an example worker that adds documents to a vectorstore, queries it, or clears it depending on the path used. It also uses Cloudflare Workers AI Embeddings.

note

If running locally, be sure to run wrangler as npx wrangler dev --remote!

name = "langchain-test"
main = "worker.ts"
compatibility_date = "2024-01-10"

[[vectorize]]
binding = "VECTORIZE_INDEX"
index_name = "langchain-test"

[ai]
binding = "AI"

// @ts-nocheck

import type {
  VectorizeIndex,
  Fetcher,
  Request,
} from "@cloudflare/workers-types";

import {
  CloudflareVectorizeStore,
  CloudflareWorkersAIEmbeddings,
} from "@langchain/cloudflare";

export interface Env {
  VECTORIZE_INDEX: VectorizeIndex;
  AI: Fetcher;
}

export default {
  async fetch(request: Request, env: Env) {
    const { pathname } = new URL(request.url);
    const embeddings = new CloudflareWorkersAIEmbeddings({
      binding: env.AI,
      model: "@cf/baai/bge-small-en-v1.5",
    });
    const store = new CloudflareVectorizeStore(embeddings, {
      index: env.VECTORIZE_INDEX,
    });
    if (pathname === "/") {
      const results = await store.similaritySearch("hello", 5);
      return Response.json(results);
    } else if (pathname === "/load") {
      // Upsertion by id is supported
      await store.addDocuments(
        [
          {
            pageContent: "hello",
            metadata: {},
          },
          {
            pageContent: "world",
            metadata: {},
          },
          {
            pageContent: "hi",
            metadata: {},
          },
        ],
        { ids: ["id1", "id2", "id3"] }
      );

      return Response.json({ success: true });
    } else if (pathname === "/clear") {
      await store.delete({ ids: ["id1", "id2", "id3"] });
      return Response.json({ success: true });
    }

    return Response.json({ error: "Not Found" }, { status: 404 });
  },
};

API Reference:

• CloudflareVectorizeStore from @langchain/cloudflare
• CloudflareWorkersAIEmbeddings from @langchain/cloudflare

You can also pass a filter parameter to filter by previously loaded metadata. See the official documentation for information on the required format.