Build a Retrieval Augmented Generation (RAG) AI
This guide will instruct you through setting up and deploying your first application with Cloudflare AI. You will build a fully-featured AI-powered application, using tools like Workers AI, Vectorize, D1, and Cloudflare Workers.
At the end of this tutorial, you will have built an AI tool that allows you to store information and query it using a Large Language Model. This pattern, known as Retrieval Augmented Generation, or RAG, is a useful project you can build by combining multiple aspects of Cloudflare's AI toolkit. You do not need to have experience working with AI tools to build this application.
- Sign up for a Cloudflare account ↗.
- Install
Node.js
↗.
Node.js version manager
Use a Node version manager like Volta ↗ or nvm ↗ to avoid permission issues and change Node.js versions. Wrangler, discussed later in this guide, requires a Node version of 16.17.0
or later.
You will also need access to Vectorize. During this tutorial, we will show how you can optionally integrate with Anthropic Claude ↗ as well. You will need an Anthropic API key ↗ to do so.
C3 (create-cloudflare-cli
) is a command-line tool designed to help you setup and deploy Workers to Cloudflare as fast as possible.
Open a terminal window and run C3 to create your Worker project:
npm create cloudflare@latest -- rag-ai-tutorial
pnpm create cloudflare@latest rag-ai-tutorial
yarn create cloudflare rag-ai-tutorial
For setup, select the following options:
- For What would you like to start with?, choose
Hello World example
. - For Which template would you like to use?, choose
Hello World Worker
. - For Which language do you want to use?, choose
JavaScript
. - For Do you want to use git for version control?, choose
Yes
. - For Do you want to deploy your application?, choose
No
(we will be making some changes before deploying).
In your project directory, C3 has generated several files.
What files did C3 create?
wrangler.json
: Your Wrangler configuration file.worker.js
(in/src
): A minimal'Hello World!'
Worker written in ES module syntax.package.json
: A minimal Node dependencies configuration file.package-lock.json
: Refer tonpm
documentation onpackage-lock.json
↗.node_modules
: Refer tonpm
documentationnode_modules
↗.
Now, move into your newly created directory:
cd rag-ai-tutorial
The Workers command-line interface, Wrangler, allows you to create, test, and deploy your Workers projects. C3 will install Wrangler in projects by default.
After you have created your first Worker, run the wrangler dev
command in the project directory to start a local server for developing your Worker. This will allow you to test your Worker locally during development.
npx wrangler dev --remote
You will now be able to go to http://localhost:8787 ↗ to see your Worker running. Any changes you make to your code will trigger a rebuild, and reloading the page will show you the up-to-date output of your Worker.
To begin using Cloudflare's AI products, you can add the ai
block to the wrangler.toml / wrangler.json
file. This will set up a binding to Cloudflare's AI models in your code that you can use to interact with the available AI models on the platform.
This example features the @cf/meta/llama-3-8b-instruct
model, which generates text.
{ "ai": { "binding": "AI" }}
[ai]binding = "AI"
Now, find the src/index.js
file. Inside the fetch
handler, you can query the AI
binding:
export default { async fetch(request, env, ctx) { const answer = await env.AI.run("@cf/meta/llama-3-8b-instruct", { messages: [{ role: "user", content: `What is the square root of 9?` }], });
return new Response(JSON.stringify(answer)); },};
By querying the LLM through the AI
binding, we can interact directly with Cloudflare AI's large language models directly in our code. In this example, we are using the @cf/meta/llama-3-8b-instruct
model, which generates text.
You can deploy your Worker using wrangler
:
npx wrangler deploy
Making a request to your Worker will now generate a text response from the LLM, and return it as a JSON object.
curl https://example.username.workers.dev
{"response":"Answer: The square root of 9 is 3."}
Embeddings allow you to add additional capabilities to the language models you can use in your Cloudflare AI projects. This is done via Vectorize, Cloudflare's vector database.
To begin using Vectorize, create a new embeddings index using wrangler
. This index will store vectors with 768 dimensions, and will use cosine similarity to determine which vectors are most similar to each other:
npx wrangler vectorize create vector-index --dimensions=768 --metric=cosine
Then, add the configuration details for your new Vectorize index to the wrangler.toml / wrangler.json
file:
{ "vectorize": [ { "binding": "VECTOR_INDEX", "index_name": "vector-index" } ]}
# ... existing wrangler configuration
[[vectorize]]binding = "VECTOR_INDEX"index_name = "vector-index"
A vector index allows you to store a collection of dimensions, which are floating point numbers used to represent your data. When you want to query the vector database, you can also convert your query into dimensions. Vectorize is designed to efficiently determine which stored vectors are most similar to your query.
To implement the searching feature, you must set up a D1 database from Cloudflare. In D1, you can store your app's data. Then, you change this data into a vector format. When someone searches and it matches the vector, you can show them the matching data.
Create a new D1 database using wrangler
:
npx wrangler d1 create database
Then, paste the configuration details output from the previous command into the wrangler.toml / wrangler.json
file:
{ "d1_databases": [ { "binding": "DB", "database_name": "database", "database_id": "abc-def-geh" } ]}
# ... existing wrangler configuration
[[d1_databases]]binding = "DB" # available in your Worker on env.DBdatabase_name = "database"database_id = "abc-def-geh" # replace this with a real database_id (UUID)
In this application, we'll create a notes
table in D1, which will allow us to store notes and later retrieve them in Vectorize. To create this table, run a SQL command using wrangler d1 execute
:
npx wrangler d1 execute database --remote --command "CREATE TABLE IF NOT EXISTS notes (id INTEGER PRIMARY KEY, text TEXT NOT NULL)"
Now, we can add a new note to our database using wrangler d1 execute
:
npx wrangler d1 execute database --remote --command "INSERT INTO notes (text) VALUES ('The best pizza topping is pepperoni')"
Before we begin creating notes, we will introduce a Cloudflare Workflow. This will allow us to define a durable workflow that can safely and robustly execute all the steps of the RAG process.
To begin, add a new [[workflows]]
block to wrangler.toml
:
# ... existing wrangler configuration
[[workflows]]name = "rag"binding = "RAG_WORKFLOW"class_name = "RAGWorkflow"
In src/index.js
, add a new class called RAGWorkflow
that extends Workflow
:
export class RAGWorkflow { async run(event, step) { await step.do('example step', async () => { console.log("Hello World!") }) }}
This class will define a single workflow step that will log "Hello World!" to the console. You can add as many steps as you need to your workflow.
On its own, this workflow will not do anything. To execute the workflow, we will call the RAG_WORKFLOW
binding, passing in any parameters that the workflow needs to properly complete. Here is an example of how we can call the workflow:
env.RAG_WORKFLOW.create({ params: { text } })
To expand on your Workers function in order to handle multiple routes, we will add hono
, a routing library for Workers. This will allow us to create a new route for adding notes to our database. Install hono
using npm
:
npm install hono
Then, import hono
into your src/index.js
file. You should also update the fetch
handler to use hono
:
import { Hono } from "hono";const app = new Hono();
app.get("/", async (c) => { const answer = await c.env.AI.run("@cf/meta/llama-3-8b-instruct", { messages: [{ role: "user", content: `What is the square root of 9?` }], });
return c.json(answer);});
export default app;
This will establish a route at the root path /
that is functionally equivalent to the previous version of your application.
Now, we can update our workflow to begin adding notes to our database, and generating the related embeddings for them.
This example features the @cf/baai/bge-base-en-v1.5
model, which can be used to create an embedding. Embeddings are stored and retrieved inside Vectorize, Cloudflare's vector database. The user query is also turned into an embedding so that it can be used for searching within Vectorize.
export class RAGWorkflow { async run(event, step) { const { text } = event.params
const record = await step.do(`create database record`, async () => { const query = "INSERT INTO notes (text) VALUES (?) RETURNING *"
const { results } = await env.DATABASE.prepare(query) .bind(text) .run()
const record = results[0] if (!record) throw new Error("Failed to create note") return record; })
const embedding = await step.do(`generate embedding`, async () => { const embeddings = await env.AI.run('@cf/baai/bge-base-en-v1.5', { text: text }) const values = embeddings.data[0] if (!values) throw new Error("Failed to generate vector embedding") return values })
await step.do(`insert vector`, async () => { return env.VECTOR_INDEX.upsert([ { id: record.id.toString(), values: embedding, } ]); }) }}
The workflow does the following things:
- Accepts a
text
parameter. - Insert a new row into the
notes
table in D1, and retrieve theid
of the new row. - Convert the
text
into a vector using theembeddings
model of the LLM binding. - Upsert the
id
andvectors
into thevector-index
index in Vectorize.
By doing this, you will create a new vector representation of the note, which can be used to retrieve the note later.
To complete the code, we will add a route that allows users to submit notes to the database. This route will parse the JSON request body, get the note
parameter, and create a new instance of the workflow, passing the parameter:
app.post('/notes', async (c) => { const { text } = await c.req.json(); if (!text) return c.text("Missing text", 400); await c.env.RAG_WORKFLOW.create({ params: { text } }) return c.text("Created note", 201);})
To complete your code, you can update the root path (/
) to query Vectorize. You will convert the query into a vector, and then use the vector-index
index to find the most similar vectors.
The topK
parameter limits the number of vectors returned by the function. For instance, providing a topK
of 1 will only return the most similar vector based on the query. Setting topK
to 5 will return the 5 most similar vectors.
Given a list of similar vectors, you can retrieve the notes that match the record IDs stored alongside those vectors. In this case, we are only retrieving a single note - but you may customize this as needed.
You can insert the text of those notes as context into the prompt for the LLM binding. This is the basis of Retrieval-Augmented Generation, or RAG: providing additional context from data outside of the LLM to enhance the text generated by the LLM.
We'll update the prompt to include the context, and to ask the LLM to use the context when responding:
import { Hono } from "hono";const app = new Hono();
// Existing post route...// app.post('/notes', async (c) => { ... })
app.get('/', async (c) => { const question = c.req.query('text') || "What is the square root of 9?"
const embeddings = await c.env.AI.run('@cf/baai/bge-base-en-v1.5', { text: question }) const vectors = embeddings.data[0]
const vectorQuery = await c.env.VECTOR_INDEX.query(vectors, { topK: 1 }); let vecId; if (vectorQuery.matches && vectorQuery.matches.length > 0 && vectorQuery.matches[0]) { vecId = vectorQuery.matches[0].id; } else { console.log("No matching vector found or vectorQuery.matches is empty"); }
let notes = [] if (vecId) { const query = `SELECT * FROM notes WHERE id = ?` const { results } = await c.env.DB.prepare(query).bind(vecId).all() if (results) notes = results.map(vec => vec.text) }
const contextMessage = notes.length ? `Context:\n${notes.map(note => `- ${note}`).join("\n")}` : ""
const systemPrompt = `When answering the question or responding, use the context provided, if it is provided and relevant.`
const { response: answer } = await c.env.AI.run( '@cf/meta/llama-3-8b-instruct', { messages: [ ...(notes.length ? [{ role: 'system', content: contextMessage }] : []), { role: 'system', content: systemPrompt }, { role: 'user', content: question } ] } )
return c.text(answer);});
app.onError((err, c) => { return c.text(err);});
export default app;
If you are working with larger documents, you have the option to use Anthropic's Claude models ↗, which have large context windows and are well-suited to RAG workflows.
To begin, install the @anthropic-ai/sdk
package:
npm install @anthropic-ai/sdk
In src/index.js
, you can update the GET /
route to check for the ANTHROPIC_API_KEY
environment variable. If it's set, we can generate text using the Anthropic SDK. If it isn't set, we'll fall back to the existing Workers AI code:
import Anthropic from '@anthropic-ai/sdk';
app.get('/', async (c) => { // ... Existing code const systemPrompt = `When answering the question or responding, use the context provided, if it is provided and relevant.`
let modelUsed: string = "" let response = null
if (c.env.ANTHROPIC_API_KEY) { const anthropic = new Anthropic({ apiKey: c.env.ANTHROPIC_API_KEY })
const model = "claude-3-5-sonnet-latest" modelUsed = model
const message = await anthropic.messages.create({ max_tokens: 1024, model, messages: [ { role: 'user', content: question } ], system: [systemPrompt, notes ? contextMessage : ''].join(" ") })
response = { response: message.content.map(content => content.text).join("\n") } } else { const model = "@cf/meta/llama-3.1-8b-instruct" modelUsed = model
response = await c.env.AI.run( model, { messages: [ ...(notes.length ? [{ role: 'system', content: contextMessage }] : []), { role: 'system', content: systemPrompt }, { role: 'user', content: question } ] } ) }
if (response) { c.header('x-model-used', modelUsed) return c.text(response.response) } else { return c.text("We were unable to generate output", 500) }})
Finally, you'll need to set the ANTHROPIC_API_KEY
environment variable in your Workers application. You can do this by using wrangler secret put
:
$ npx wrangler secret put ANTHROPIC_API_KEY
If you no longer need a note, you can delete it from the database. Any time that you delete a note, you will also need to delete the corresponding vector from Vectorize. You can implement this by building a DELETE /notes/:id
route in your src/index.js
file:
app.delete("/notes/:id", async (c) => { const { id } = c.req.param();
const query = `DELETE FROM notes WHERE id = ?`; await c.env.DB.prepare(query).bind(id).run();
await c.env.VECTOR_INDEX.deleteByIds([id]);
return c.status(204);});
For large pieces of text, it is recommended to split the text into smaller chunks. This allows LLMs to more effectively gather relevant context, without needing to retrieve large pieces of text.
To implement this, we'll add a new NPM package to our project, `@langchain/textsplitters':
npm install @cloudflare/textsplitters
The RecursiveCharacterTextSplitter
class provided by this package will split the text into smaller chunks. It can be customized to your liking, but the default config works in most cases:
import { RecursiveCharacterTextSplitter } from "@langchain/textsplitters";
const text = "Some long piece of text...";
const splitter = new RecursiveCharacterTextSplitter({ // These can be customized to change the chunking size // chunkSize: 1000, // chunkOverlap: 200,});
const output = await splitter.createDocuments([text]);console.log(output) // [{ pageContent: 'Some long piece of text...' }]
To use this splitter, we'll update the workflow to split the text into smaller chunks. We'll then iterate over the chunks and run the rest of the workflow for each chunk of text:
export class RAGWorkflow { async run(event, step) { const env = this.env const { text } = event.payload; let texts = await step.do('split text', async () => { const splitter = new RecursiveCharacterTextSplitter(); const output = await splitter.createDocuments([text]); return output.map(doc => doc.pageContent); })
console.log("RecursiveCharacterTextSplitter generated ${texts.length} chunks")
for (const index in texts) { const text = texts[index] const record = await step.do(`create database record: ${index}/${texts.length}`, async () => { const query = "INSERT INTO notes (text) VALUES (?) RETURNING *"
const { results } = await env.DATABASE.prepare(query) .bind(text) .run()
const record = results[0] if (!record) throw new Error("Failed to create note") return record; })
const embedding = await step.do(`generate embedding: ${index}/${texts.length}`, async () => { const embeddings = await env.AI.run('@cf/baai/bge-base-en-v1.5', { text: text }) const values = embeddings.data[0] if (!values) throw new Error("Failed to generate vector embedding") return values })
await step.do(`insert vector: ${index}/${texts.length}`, async () => { return env.VECTOR_INDEX.upsert([ { id: record.id.toString(), values: embedding, } ]); }) } }}
Now, when large pieces of text are submitted to the /notes
endpoint, they will be split into smaller chunks, and each chunk will be processed by the workflow.
If you did not deploy your Worker during step 1, deploy your Worker via Wrangler, to a *.workers.dev
subdomain, or a Custom Domain, if you have one configured. If you have not configured any subdomain or domain, Wrangler will prompt you during the publish process to set one up.
npx wrangler deploy
Preview your Worker at <YOUR_WORKER>.<YOUR_SUBDOMAIN>.workers.dev
.
A full version of this codebase is available on GitHub. It includes a frontend UI for querying, adding, and deleting notes, as well as a backend API for interacting with the database and vector index. You can find it here: github.com/kristianfreeman/cloudflare-retrieval-augmented-generation-example ↗.
To do more:
- Explore the reference diagram for a Retrieval Augmented Generation (RAG) Architecture.
- Review Cloudflare's AI documentation.
- Review Tutorials to build projects on Workers.
- Explore Examples to experiment with copy and paste Worker code.
- Understand how Workers works in Reference.
- Learn about Workers features and functionality in Platform.
- Set up Wrangler to programmatically create, test, and deploy your Worker projects.