Build your own documentation chatbot with ChatGPT and EdgeDB

For additional context, check out our blog post about why and how we use ChatGPT via embeddings to create our “Ask AI” bot which answers questions related to the EdgeDB docs.

In this tutorial we’re going to build a documentation chatbot with Next.js, OpenAI, and EdgeDB.

Warning: This project makes calls to OpenAI’s APIs. At the time of publication, new users are granted $5 in API credits to use within the first three months after registration. Trials are granted per phone number, not per account. If you exhaust your trial credits or if your three months lapse, you will need to switch to a paid account in order to build the tutorial project.

How it works​

tl;dr- Training a language model is hard, but using embeddings to give it access to information beyond what it’s trained on is easy… so we will do that! Now, skip ahead to get started building or read on for more detail.

Our chatbot is backed by OpenAI’s ChatGPT. ChatGPT is an advanced large language model (LLM) that uses machine learning algorithms to generate human-like responses based on the input it’s given.

There are two options when integrating ChatGPT and language models in general: fine-tuning the model or using embeddings. Fine-tuning produces the best result, but it needs more of everything: more money, more time, more resources, and more training data. That’s why many people and businesses use embeddings instead to provide additional context to an existing language model.

Embeddings are a way to convert words, phrases, or other types of data into a numerical form that a computer can do math with. All of this is built on top of the foundation of natural language processing (NLP) which allows computers to fake an understanding of human language. In the context of NLP, word embeddings are used to transform words into vectors. These vectors define a word’s position in space where the computer sorts them based on their syntactic and semantic similarity. For instance, synonyms are closer to each other, and words that often appear in similar contexts are grouped together.

When using embeddings we are not training the language model. Instead we’re creating embedding vectors for every piece of documentation which will later help us find which piece of documentation likely answers a user’s question. When a user asks a question, we create a new embedding for that question and compare it against the embeddings generated from our documentation to find the most similar embeddings. The answer is generated using the content that corresponds to these similar embeddings.

With that out of the way, let’s walk through how the pieces fit together.

Implementation overview​

Broadly, the app does two things: it generates embeddings from documentation, and it uses those embeddings to answer user questions. The first is triggered manually in this implementation. We’ll want to trigger it whenever the documentation is updated. The second is triggered automatically when the user asks a question.

Embedding generation requires two steps:

1. create embeddings for each section using OpenAI’s embeddings API
2. store the embeddings data in EdgeDB using pgvector

Each time a user asks a question, our app will:

1. query the database for the documentation sections most relevant to the question using a similarity function
2. inject these sections as a context into the prompt — together with user’s question — and submit this request to OpenAI
3. stream the OpenAI response back to the user in real time

Prerequisites​

This tutorial assumes you have Node.js installed. If you don’t, please install it before continuing.

The build requires other software too, but we’ll help you install it as part of the tutorial.

Initial setup​

Let’s start by scaffolding our app with the Next.js create-next-app tool. Run this wherever you would like to create the new directory for this project.

npx create-next-app --typescript docs-chatbot

Need to install the following packages:
  create-next-app@13.4.12
Ok to proceed? (y) y
✔ Would you like to use ESLint? … No / Yes
✔ Would you like to use Tailwind CSS? … No / Yes
✔ Would you like to use `src/` directory? … No / Yes
✔ Would you like to use App Router? (recommended) … No / Yes
✔ Would you like to customize the default import alias? … No / Yes
Creating a new Next.js app in /<path>/<to>/<project>/docs-chatbot.

Choose “Yes” for all questions except “Would you like to use src/\ directory?” and “Would you like to customize the default import alias?”

Once bootstrapping is complete, you should see a success message:

Success! Created docs-chatbot at /<path>/<to>/<project>/docs-chatbot

Change into the new directory so we can get started!

cd docs-chatbot

Let’s make two changes to the tsconfig.json generated by create-next-app. Change the target to "es6" because we will use some data structures that are only available in ES6. Update the compilerOptions object by setting the baseUrl property to the root with "baseUrl": ".". Later when we add modules to the root of the project, this will make it easier to import them.

{
  "compilerOptions": {
    "target": "es6",
    "lib": ["dom", "dom.iterable", "esnext"],
    "allowJs": true,
    "skipLibCheck": true,
    "strict": true,
    "forceConsistentCasingInFileNames": true,
    "noEmit": true,
    "esModuleInterop": true,
    "module": "esnext",
    "moduleResolution": "bundler",
    "resolveJsonModule": true,
    "isolatedModules": true,
    "jsx": "preserve",
    "incremental": true,
    "plugins": [
      {
        "name": "next"
      }
    ],
    "paths": {
      "@/*": ["./*"]
    },
    "baseUrl": "."
  },
  "include": ["next-env.d.ts", "**/*.ts", "**/*.tsx", ".next/types/**/*.ts"],
  "exclude": ["node_modules"]
}

Now, we’ll create an instance of EdgeDB for our project, but first, we need to install EdgeDB!

Install the EdgeDB CLI​

If you already have EdgeDB installed, you can skip to creating an instance.

Before we can create an instance for our project, we need to install the EdgeDB CLI. On Linux or MacOS, run the following in your terminal and follow the on-screen instructions:

curl --proto '=https' --tlsv1.2 -sSf https://sh.edgedb.com | sh

Windows Powershell users can use this command:

iwr https://ps1.edgedb.com -useb | iex

For other installation scenarios, see the “Additional installation methods” section of our “Install” page.

Create a local EdgeDB instance​

To create our instance, let’s initialize our project as an EdgeDB project. Run the following in the root of the project:

edgedb project init

No `edgedb.toml` found in `/<path>/<to>/<project>/docs-chatbot` or above

Do you want to initialize a new project? [Y/n]
> Y

Specify the name of EdgeDB instance to use with this project [default: docs_chatbot]:
> docs_chatbot

Checking EdgeDB versions...
Specify the version of EdgeDB to use with this project
[default: 3.2]:
> 3.2

The CLI should set up an EdgeDB project, an instance, and a database within that instance.

• Confirm project creation by checking for an edgedb.toml file and a dbschema directory in the project root.

• Confirm the instance is running with the edgedb instance list command. Search for the name of the instance you’ve just created (docs_chatbot if you’re following along) and check the status. (Don’t worry if the status is “inactive”; the status will change to “running” automatically when you connect to the instance.)

• Confirm you can connect to the created instance by running edgedb in the terminal to connect to it via REPL or by running edgedb ui to connect using the UI.

Configure the environment​

Create a .env.local file in the root of your new Next.js project.

touch .env.local

We’re going to add a couple of variables to that file to configure the EdgeDB client. We’ll need to run a command on our new instance to get the value for one of those. Since we’ll be using the Edge runtime in our Next.js project, the edgedb-js client won’t be able to access the Node.js filesystem APIs it usually uses to automatically find your instance, so we need to provide the DSN for the instance instead. To get that, run this command:

edgedb instance credentials --insecure-dsn

Copy what it logs out. Open the .env.local file in your text editor and add this to it:

EDGEDB_DSN=<your-dsn>
EDGEDB_CLIENT_TLS_SECURITY="insecure"

Replace <your-dsn> with the value you copied earlier.

We’re going to be using the EdgeDB HTTP client a bit later to connect to our database, but it requires a trusted TLS/SSL certificate. Local development instances use self signed certificates, and using HTTPS with these certificates will result in an error. To work around this error, we allow the client to ignore TLS by setting the EDGEDB_CLIENT_TLS_SECURITY variable to "insecure". Bear in mind that this is only for local development, and you should always use TLS in production.

We need to set one more environment variable, but first we have to get an API key.

Prepare the OpenAI API client​

We need an API key from OpenAI in order to make the calls we need to make this app work. To get one:

1. Log in or sign up to the OpenAI platform.
2. Create new secret key.

Warning: Don’t forget you may need to start a paid account if you do not have any API free trial credits remaining.

Copy the new key. Re-open your .env.local file and add it like this:

EDGEDB_DSN=<your-dsn>
EDGEDB_CLIENT_TLS_SECURITY="insecure"
OPENAI_API_KEY="<your-openai-api-key>"

Instead of <your-openai-api-key>, paste in the key you just created.

While we’re here, let’s get that key ready to be used. We will be making calls to the OpenAI API. We’ll create a utils module and export a function from it that initializes an OpenAI API client. We can import and call the function to create a new client anywhere we need to make OpenAI API calls.

// utils.ts in the project root
import OpenAI from "openai";

export function initOpenAIClient() {
  if (!process.env.OPENAI_API_KEY)
    throw new Error("Missing environment variable OPENAI_API_KEY");

  return new OpenAI({
    apiKey: process.env.OPENAI_API_KEY,
  });
}

It’s pretty simple. It makes sure the API key was provided in the environment variable and returns a new API client initialized with that key.

Now, let’s create error messages we will use in a couple of places if these API calls go wrong. Create a file app/constants.ts and fill it with this:

export const errors = {
  flagged: `OpenAI has declined to answer your question due to their usage-policies. Please try another question.`,
  default: "There was an error processing your request. Please try again.",
};

This exports an object errors with a couple of error messages.

Now, let’s get the documentation ready!

Put the documentation in place​

For this project, we will be using documentation written as Markdown files since they are straightforward for OpenAI’s language models to use.

Create a docs folder in the root of the project. Here we will place our Markdown documentation files. You can grab the files we use from the example project’s GitHub repo or add your own. (If you use your own, you may also want to adjust the system message we send to OpenAI later.)

Note: On using formats other than Markdown

We could opt to use other simple formats like plain text files or more complex ones like HTML. Since more complex formats can include additional data beyond what we want the language model to consume (like HTML’s tags and their attributes), we may first want to clean those files and extract the content before sending it to OpenAI. (We can write our own logic for this or use libraries that are available online for conversion, to Markdown for example.)

It’s possible to use more complex formats without cleaning them, but then we’re paying for extra tokens that don’t improve the answers our chatbot will give users.

Note: On longer documentation sections

In this tutorial project, our documentation pages are short, but in practice, documentation files can get quite long and may need to be split into multiple sections because of the LLM’s token limit. LLMs divide text into tokens. For English text, 1 token is approximately 4 characters or 0.75 words. LLMs have limits on the number of tokens they can receive and send back.

One approach to mitigate this is to parse your documentation files and create new sections every time you encounter a header. If you use this approach, consider section lengths when writing your documentation. If you find a section is too long, consider ways you might break it up with additional headings. This will probably make it easier to read for your users too!

To generate embeddings, we will use the text-embedding-ada-002 model. Its input token limit is 8,191 tokens. Later, when answering a user’s questions we will use the chat completions model pt-3.5-turbo. Its token limit is 4,096 tokens. This limit covers not only our input, but also the API’s response.

Later, when we send the user’s question, we will also send related sections from our documentation as part of the input to the chat completions API. This is why it’s important to keep our sections short: we want to leave enough space for the answer.

If the related sections are too long and, together with the user’s question, exceed the 4,096 token limit, we will get an error back from OpenAI. If the length of the question and related sections are too close to the token limit but not over it, the API will send an answer, but the answer will be cut off when the limit is reached.

We want to avoid either of these outcomes by making sure we always have enough token headroom for all the input and the LLM’s response. That’s why we will later set 1,500 tokens as the maximum number of tokens we will use for our related sections, and it’s also why it’s important that sections be relatively short.

If your application has longer documentation files, make sure to figure out a strategy for splitting those before you generate your embeddings.

Create the schema to store embeddings​

To be able to store data in the database, we have to create its schema first. We want to make the schema as simple as possible and store only the relevant data. We need to store the section’s embeddings, content, and the number of tokens. The embeddings allow us to match content to questions. The content gives us context to feed to the LLM. We will need the token count later when calculating how many related sections fit inside the prompt context while staying under the model’s token limit.

Open the empty schema file that was generated when we initialized the EdgeDB project (located at dbschema/default.esdl from the project directory). We’ll walk through what we’ll add to it, one step at a time. First, add this at the top of the file (above module default {):

# dbschema/default.esdl

using extension pgvector;
module default {
  # Schema will go here
}

We are able to store embeddings and find similar embeddings in the EdgeDB database because of the pgvector extension. In order to use it in our schema, we have to activate the ext::pgvector module with using extension pgvector at the beginning of the schema file. This module gives us access to the ext::pgvector::vector data type as well as few similarity functions and indexes we can use later to retrieve embeddings. Read our pgvector documentation for more details on the extension.

Just below that, we can start building our module by creating a new scalar type.

# dbschema/default.esdl

using extension pgvector;
module default {
  scalar type OpenAIEmbedding extending
    ext::pgvector::vector<1536>;

  type Section {
    # We will build this out next
  }
}

With the extension active, we may now add properties to our object types using the included ext::pgvector::vector data type. However, in order to be able to use indexes, the vectors in question need to be a of a fixed length. This can be achieved by creating a custom scalar extending the vector and specifying the desired length. OpenAI embeddings have length of 1,536, so that’s what we use in our schema for this custom scalar.

Now, the Section type:

# dbschema/default.esdl

using extension pgvector;
module default {
  scalar type OpenAIEmbedding extending
    ext::pgvector::vector<1536>;

  type Section {
    required content: str;
    required tokens: int16;
    required embedding: OpenAIEmbedding;

    index ext::pgvector::ivfflat_cosine(lists := 1)
      on (.embedding);
  }
}

The Section contains properties to store the content, a count of tokens, and the embedding, which is of the custom scalar type we created in the previous step.

We’ve also added an index inside the Section type to speed up queries. In order for this to work properly, the index should correspond to the cosine_similarity function we’re going to use to find sections related to the user’s question. That corresponding index is ivfflat_cosine.

We are using the value 1 for the lists parameter because we will have very few items in our database — three, to be exact 😅. Best practice is to use the number of objects divided by 1,000 for up to 1,000,000 objects.

In our case indexing does not have much impact, but if you plan to store and query a large number of entries, you’ll see performance gains by adding this index.

Put that all together, and your entire schema file should look like this:

# dbschema/default.esdl

using extension pgvector;

module default {
  scalar type OpenAIEmbedding extending
    ext::pgvector::vector<1536>;

  type Section {
    required content: str;
    required tokens: int16;
    required embedding: OpenAIEmbedding;

    index ext::pgvector::ivfflat_cosine(lists := 1)
      on (.embedding);
  }
}

We apply this schema by creating and running a migration.

edgedb migration create
edgedb migrate

Note: In this tutorial we will regenerate all embeddings every time we run the embeddings generation script, wiping all data and saving new Section objects for all of the documentation. This might be a reasonable approach if you don’t have much documentation, but if you have a lot of documentation, you may want a more sophisticated approach that operates on only documentation sections which have changed.

You can achieve this by saving content checksums and a unique identifier for each section — in our production implementation, we use section paths — as part of your Section objects. The next time you run generation, compare the section’s current checksum with the one you stored in the database, finding it by its unique identifier. You don’t need to generate embeddings and update the database for a given section unless the two checksums are different indicating something has changed.

If you decide to go this route, here’s one way you could modify your schema to support this:

# dbschema/default.esdl
# This is only necessary to support partial regeneration
# We will not use it for this tutorial.

type Section {
  required path: str {
    constraint exclusive;
  }
  required checksum: str;
  # The rest of the Section type
}

You’ll also need to store your unique identifier, calculate and compare checksums, and update objects conditionally based on the outcome of those comparisons.

Create and store embeddings​

Before we can script the creation of embeddings, we need to install some libraries that will help us.

npm install openai edgedb
npm install \
  @edgedb/generate \
  gpt-tokenizer \
  dotenv \
  tsx \
  --save-dev

The @edgedb/generate package provides a set of code generation tools that are useful when developing an EdgeDB-backed applications with TypeScript/JavaScript. We’re going to write queries using our query builder, but before we can, we need to run the query builder generator.

npx @edgedb/generate edgeql-js

Answer “y” when asked about adding the query builder to .gitignore.

This generator gives us a code-first way to write fully-typed EdgeQL queries with TypeScript. After running the generator, you should see a new edgeql-js folder inside dbschema.

Finally, we’re ready to create embeddings for all sections and store them in the database we created earlier. Let’s make a generate-embeddings.ts file inside the project root.

touch generate-embeddings.ts

Let’s look at the script’s skeleton and get an understanding of the flow of tasks we need to perform.

Rather than trying to build this incrementally as we go, you may just want to read through to understand all the code. We’ll put the entire script together at the end of the section, and you can copy/paste that into your file.

// generate-embeddings.ts

import { promises as fs } from "fs";
import { join } from "path";
import dotenv from "dotenv";
import { encode } from "gpt-tokenizer";
import * as edgedb from "edgedb";
import e from "dbschema/edgeql-js";
import { initOpenAIClient } from "./utils";

dotenv.config({ path: ".env.local" });

const openai = initOpenAIClient();

interface Section {
  id?: string;
  tokens: number;
  content: string;
  embedding: number[];
}

async function walk(dir: string): Promise<string[]> {
  // …
}

async function prepareSectionsData(
  sectionPaths: string[]
): Promise<Section[]> {
  // …
}

async function storeEmbeddings() {
  // …
}

(async function main() {
  await storeEmbeddings();
})();

At the top are all imports we will need throughout the file. The second to last import is the query builder we generated earlier, and the last one is the function that initializes our OpenAI API client.

After the imports, we use the dotenv library to import environment variables from the .env.local file.

Then, we initialize our OpenAI API client by calling initOpenAIClient.

Next, we define a Section TypeScript interface that corresponds to the Section type we have defined in the schema.

Then we have a few function definitions:

• walk and prepareSectionsData will be called from inside storeEmbeddings. walk returns an array of all documentation page paths relative to the project root. prepareSectionsData takes care of preparing the Section objects we will insert into the database and returns those as an array.
• storeEmbeddings coordinates everything.

To finish the script, we await a call to our coordinating function which kicks off everything else as needed.

Getting section paths​

In order to get the sections’ content, we first need to know where the files are that need to be read. The walk function finds them for us and returns all the paths. It builds an array of all paths relative to the project root.

// generate-embeddings.ts

// …
async function walk(dir: string): Promise<string[]> {
  const entries = await fs.readdir(dir, { withFileTypes: true });

  return (
    await Promise.all(
      entries.map((entry) => {
        const path = join(dir, entry.name);
        if (entry.isFile()) return [path];
        else if (entry.isDirectory()) return walk(path);
        return [];
      })
    )
  ).flat();
}
// …

The output it produces looks like this:

[
  'docs/edgeql/design-goals.md',
  'docs/edgeql/overview.md',
  'docs/edgeql/try-edgeql.md',
]

Preparing the Section objects​

This function will be responsible for collecting the data we need for each Section object we will store, including making the OpenAI API calls to generate the embeddings. Let’s walk through it one piece at a time.

// generate-embeddings.ts

// …
async function prepareSectionsData(
  sectionPaths: string[]
): Promise<Section[]> {
  const contents: string[] = [];
  const sections: Section[] = [];

  for (const path of sectionPaths) {
    const content = await fs.readFile(path, "utf8");
    // OpenAI recommends replacing newlines with spaces for best results
    // when generating embeddings
    const contentTrimmed = content.replace(/\n/g, " ");
    contents.push(contentTrimmed);
    sections.push({
      content,
      tokens: encode(content).length
      embedding: [],
    });
  }
  // The rest of the function
}
// …

We start with a parameter: an array of section paths. We create a couple of empty arrays for storing information about our sections (which will later become Section objects in the database) and their contents. We iterate through the paths, loading each file to get its content.

In the database we will save the content as is, but when calling the embedding API, OpenAI suggests that all newlines should be replaced with a single space for the best results. contentTrimmed is the content with newlines replaced. We push that onto our contents array and the un-trimmed content onto sections, along with a token count (obtained by calling the encode function imported from gpt-tokenizer) and an empty array we will later replace with the actual embeddings.

Onto the next bit!

// generate-embeddings.ts

// …
async function prepareSectionsData(
  sectionPaths: string[]
): Promise<Section[]> {
  // Part we just talked about

  const embeddingResponse = await openai.embeddings.create({
    model: "text-embedding-ada-002",
    input: contents,
  });

  // The rest
}
// …

Now, we generate embeddings from the content. We need to be careful about how we approach the API calls to generate the embeddings since they could have a big impact on how long generation takes, especially as your documentation grows. The simplest solution would be to make a single request to the API for each section, but in the case of EdgeDB’s documentation, which has around 3,000 pages, this would take about half an hour.

Since OpenAI’s embeddings API can take not only a single string but also an array of strings, we can leverage this to batch up all our content and generate the embeddings with a single request! You can see that single API call when we set embeddingResponse to the result of the call to openai.embeddings.create, specifying the model and passing the entire array of contents.

One downside to this one-shot embedding generation approach is that we do not get back token counts with the result where we would generating embeddings for only a single string. Token counts are important because they determine how many relevant sections we can send along with our input to the chat completions API — the one that answers the user’s question — and still be within the model’s token limit. To stay within the limit, we need to know how many tokens each section has. Since we don’t get them back on a batched embedding generation, we used the gpt-tokenizer library’s encode function earlier to count them ourselves.

Now, it’s time to put those embeddings into our section objects by iterating through the response data.

// generate-embeddings.ts

// …
async function prepareSectionsData(
  sectionPaths: string[]
): Promise<Section[]> {
  // The stuff we already talked about

  embeddingResponse.data.forEach((item, i) => {
    sections[i].embedding = item.embedding;
  });

  return sections;
}
// …

We iterate through all the embeddings we got back, adding the embedding to its respective section. This final piece of data makes the section fully ready to store in the database, so we can now return the fully-formed sections from the function.

Here’s the entire function assembled:

// generate-embeddings.ts

// …
async function prepareSectionsData(
  sectionPaths: string[]
): Promise<Section[]> {
  const contents: string[] = [];
  const sections: Section[] = [];

  for (const path of sectionPaths) {
    const content = await fs.readFile(path, "utf8");
    // OpenAI recommends replacing newlines with spaces for best results
    // when generating embeddings
    const contentTrimmed = content.replace(/\n/g, " ");
    contents.push(contentTrimmed);
    sections.push({
      content,
      tokens: encode(content).length
      embedding: [],
    });
  }

  const embeddingResponse = await openai.embeddings.create({
    model: "text-embedding-ada-002",
    input: contents,
  });

  embeddingResponse.data.forEach((item, i) => {
    sections[i].embedding = item.embedding;
  });

  return sections;
}
// …

This is not the only approach to keeping track of tokens. We could choose not to save token counts in the database and to instead count section tokens later on the client after we find the relevant sections.

Now that we have sections ready to be stored in the database, let’s tie everything together with the storeEmbeddings function.

Storing the Section objects​

Again, we’ll break the storeEmbeddings function apart and walk through it.

// generate-embeddings.ts

// …
async function storeEmbeddings() {
  const client = edgedb.createClient();

  const sectionPaths = await walk("docs");

  console.log(`Discovered ${sectionPaths.length} sections`);

  const sections = await prepareSectionsData(sectionPaths);

  // The rest of the function
}
// …

We create our EdgeDB client and get our documentation paths by calling walk. We also log out some debug information showing how many sections were discovered. Then, we prep our Section objects by calling the prepareSectionsData function we just walked through and passing in the documentation paths we got back from walk.

Next, we’ll store this data.

// generate-embeddings.ts

// …
async function storeEmbeddings() {
  // The parts we just talked about

  // Delete old data from the DB.
  await e.delete(e.Section).run(client);

  // Bulk-insert all data into EdgeDB database.
  const query = e.params({ sections: e.json }, ({ sections }) => {
    return e.for(e.json_array_unpack(sections), (section) => {
      return e.insert(e.Section, {
        content: e.cast(e.str, section.content),
        tokens: e.cast(e.int16, section.tokens),
        embedding: e.cast(e.OpenAIEmbedding, section.embedding),
      });
    });
  });

  await query.run(client, { sections });
  console.log("Embedding generation complete");
}
// …

The comments do a good job of explaining here, but let’s go into a little more detail. First, we build and run a query that deletes all Section objects currently in the database. Then, we build another query that will insert the new Section data we just prepared. We await a call to that query’s run method, passing in the sections we just prepared.

Here’s what the whole function looks like:

// generate-embeddings.ts

// …
async function storeEmbeddings() {
  const client = edgedb.createClient();

  const sectionPaths = await walk("docs");

  console.log(`Discovered ${sectionPaths.length} sections`);

  const sections = await prepareSectionsData(sectionPaths);

  // Delete old data from the DB.
  await e.delete(e.Section).run(client);

  // Bulk-insert all data into EdgeDB database.
  const query = e.params({ sections: e.json }, ({ sections }) => {
    return e.for(e.json_array_unpack(sections), (section) => {
      return e.insert(e.Section, {
        content: e.cast(e.str, section.content),
        tokens: e.cast(e.int16, section.tokens),
        embedding: e.cast(e.OpenAIEmbedding, section.embedding),
      });
    });
  });

  await query.run(client, { sections });
  console.log("Embedding generation complete");
}
// …

Putting it all together​

Here’s the entire embeddings generation script. Copy and paste the whole thing into your generate-embeddings.ts file.

// generate-embeddings.ts

import { promises as fs } from "fs";
import { join } from "path";
import dotenv from "dotenv";
import { encode } from "gpt-tokenizer";
import * as edgedb from "edgedb";
import e from "dbschema/edgeql-js";
import { initOpenAIClient } from "@/utils";

dotenv.config({ path: ".env.local" });

const openai = initOpenAIClient();

interface Section {
  id?: string;
  tokens: number;
  content: string;
  embedding: number[];
}

async function walk(dir: string): Promise<string[]> {
  const entries = await fs.readdir(dir, { withFileTypes: true });

  return (
    await Promise.all(
      entries.map((entry) => {
        const path = join(dir, entry.name);
        if (entry.isFile()) return [path];
        else if (entry.isDirectory()) return walk(path);
        return [];
      })
    )
  ).flat();
}

async function prepareSectionsData(
  sectionPaths: string[]
): Promise<Section[]> {
  const contents: string[] = [];
  const sections: Section[] = [];

  for (const path of sectionPaths) {
    const content = await fs.readFile(path, "utf8");
    // OpenAI recommends replacing newlines with spaces for best results
    // when generating embeddings
    const contentTrimmed = content.replace(/\n/g, " ");
    contents.push(contentTrimmed);
    sections.push({
      content,
      tokens: encode(content).length,
      embedding: [],
    });
  }

  const embeddingResponse = await openai.embeddings.create({
    model: "text-embedding-ada-002",
    input: contents,
  });

  embeddingResponse.data.forEach((item, i) => {
    sections[i].embedding = item.embedding;
  });

  return sections;
}

async function storeEmbeddings() {
  const client = edgedb.createClient();

  const sectionPaths = await walk("docs");

  console.log(`Discovered ${sectionPaths.length} sections`);

  const sections = await prepareSectionsData(sectionPaths);

  // Delete old data from the DB.
  await e.delete(e.Section).run(client);

  // Bulk-insert all data into EdgeDB database.
  const query = e.params({ sections: e.json }, ({ sections }) => {
    return e.for(e.json_array_unpack(sections), (section) => {
      return e.insert(e.Section, {
        content: e.cast(e.str, section.content),
        tokens: e.cast(e.int16, section.tokens),
        embedding: e.cast(e.OpenAIEmbedding, section.embedding),
      });
    });
  });

  await query.run(client, { sections });
  console.log("Embedding generation complete");
}

(async function main() {
  await storeEmbeddings();
})();

Running the script​

Let’s add a script to package.json that will invoke and execute generate-embeddings.ts. Add scripts.embeddings below:

{
  "name": "docs-chatbot",
  "version": "0.1.0",
  "private": true,
  "scripts": {
    "dev": "next dev",
    "build": "next build",
    "start": "next start",
    "lint": "next lint"
    "lint": "next lint",
    "embeddings": "tsx generate-embeddings.ts"
  },
  "dependencies": {
    "edgedb": "^1.3.5",
    "next": "^13.4.19",
    "openai": "^4.0.1",
    "react": "18.2.0",
    "react-dom": "18.2.0",
    "typescript": "5.1.6"
  },
  "devDependencies": {
    "@edgedb/generate": "^0.3.3",
    "@types/node": "20.4.8",
    "@types/react": "18.2.18",
    "@types/react-dom": "18.2.7",
    "autoprefixer": "10.4.14",
    "dotenv": "^16.3.1",
    "eslint": "8.46.0",
    "eslint-config-next": "13.4.13",
    "gpt-tokenizer": "^2.1.1",
    "postcss": "8.4.27",
    "tailwindcss": "3.3.3",
    "tsx": "^3.12.7"
  }
}

Now we can invoke the generate-embeddings.ts script from our terminal using a simple command:

npm run embeddings

After the script finishes, open the EdgeDB UI.

edgedb ui

Open your “edgedb” database and switch to the Data Explorer tab. You should see that the database has been updated with the embeddings and other relevant data.

Answering user questions​

Now that we have the content’s embeddings stored, we can start working on the handler for user questions. The user will submit a question to our server, and the handler will send them an answer back. We will define a route and an HTTP request handler for this task. Thanks to the power of Next.js, we can do all of this within our project using a route handler.

As we write our handler, one important consideration is that answers can be quite long. We could wait on the server side to get the whole answer from OpenAI and then send it to the client, but that would feel slow to the user. OpenAI supports streaming, so instead we can send the answer to the client in chunks, as they arrive to the server. With this approach, the user doesn’t have to wait for the entire response before they start getting feedback and our API seems faster.

In order to stream responses, we will use the browser’s server-sent events (SSE) API. Server-sent events enable a client to receive automatic updates from a server via an HTTP connection, and describes how the server maintains data transmissions to a client once an initial client connection has been established. The client sends a request and with that request initiates a connection with the server. The server then sends data back to the client in chunks until all of the data is sent, at which point it closes the connection.

Next.js route handler​

When using Next.js’s App Router, route handlers should be written inside an app/api folder. Every route should have its own folder within that, and the handlers should be defined inside a route.ts file inside the route’s folder.

Let’s create a new folder for the answer generation route inside app/api.

mkdir app/api && cd app/api
mkdir generate-answer && touch generate-answer/route.ts

We also need to install the common-tags NPM package (and its corresponding types package) which gives us some useful template tags that we will use later when we create the prompt from user’s question and related sections.

npm install common-tags
npm install @types/common-tags --save-dev

Let’s talk briefly about runtimes. In the context of Next.js, “runtime” refers to the set of libraries, APIs, and general functionality available to your code during execution. Next.js supports Node.js and Edge runtimes. (The “Edge” runtime is coincidentally named but is not related to EdgeDB.)

Streaming is supported within both runtimes, but the implementation is a bit simpler when using Edge, so that’s what we will use here. The Edge runtime is based on Web APIs. It has very low latency thanks to its minimal use of resources, but the downside is that it doesn’t support native Node.js APIs.

We’ll start by importing the modules we will need in the handler and writing some configuration.

Like before, you may want to read along for understanding and copy/paste the completed route at the end of this section.

// app/api/generate-answer/route.ts

import { stripIndents, oneLineTrim } from "common-tags";
import * as edgedb from "edgedb";
import e from "dbschema/edgeql-js";
import { errors } from "../../constants";
import { initOpenAIClient } from "@/utils";

export const runtime = "edge";

const openai = initOpenAIClient();

const client = edgedb.createHttpClient();

export async function POST(req: Request) {
    // …
}

// other functions that are called inside POST handler

The first imports are templates from the common-tags library we installed earlier. Then, we import the EdgeDB binding. The third import is the query builder we described previously. We also import our errors and our OpenAI API client initializer function.

By exporting runtime, we override the Next.js default for this handler so that Next.js will use the Edge runtime instead of the default Node.js runtime.

We’re ready now to write the handler function for HTTP POST requests. To do this in Next.js, you export a function named for the request method you want it to handle.

Our POST handler calls other functions that we won’t define just yet, but we’ll circle back to them later.

// app/api/generate-answer/route.ts

// …

export async function POST(req: Request) {
  try {
    const { query } = await req.json();
    const sanitizedQuery = query.trim();

    const flagged = await isQueryFlagged(query);

    if (flagged) throw new Error(errors.flagged);

    const embedding = await getEmbedding(query);

    const context = await getContext(embedding);

    const prompt = createFullPrompt(sanitizedQuery, context);

    const answer = await getOpenAiAnswer(prompt);

    return new Response(answer.body, {
      headers: {
        "Content-Type": "text/event-stream",
      },
    });
  } catch (error: any) {
    console.error(error);

    const uiError = error.message || errors.default;

    return new Response(uiError, {
      status: 500,
      headers: { "Content-Type": "application/json" },
    });
  }
}

Our handler will run the user’s question through a few different steps as we build toward an answer.

1. We check that the query complies with the OpenAI’s usage policies, which means that it should not include any hateful, harassing, or violent content. This is handled by our isQueryFlagged function.
2. If the query fails, we throw. If it passes, we generate embeddings for it using the OpenAI embedding API. This is handled by our getEmbedding function.
3. We get related documentation sections from the EdgeDB database. This is handled by getContext.
4. We create the full prompt as our input to the chat completions API by combining the question, related documentation sections, and a system message.

The system message is a general instruction to the language model that it should follow when answering any question.

With the input fully prepared, we call the chat completions API using the previously generated prompt, and we stream the response we get from OpenAI to the user. In order to use streaming we need to provide the appropriate content-type header: "text/event-stream". (You can see that in the options object passed to the Response constructor.)

To keep things simple, we’ve wrapped most of these in a single try/catch block. If any error occurs we send the error message to the user with status 500. In practice, you may want to split this up and respond with different status codes based on the outcome. For example, in the case the moderation request returns an error, you may want to send back a 400 response status (“Bad Request”) instead of a 500 (“Internal Server Error”).

Now that you can see broadly what we’re doing in this handler, let’s dig into each of the functions we’ve called in it.

Moderation request​

Let’s look at our moderation request function: isQueryFlagged. We will use the openai.moderations.create method.

// app/api/generate-answer/route.ts

async function isQueryFlagged(query: string) {
  const moderation = await openai.moderations.create({
    input: query,
  });

  const [{ flagged }] = moderation.results;

  return flagged;
}

The function is pretty straightforward: it takes the question (the query parameter), fires off a moderation request to the API, unpacks flagged from the results, and returns it.

If the API finds an issue with the user’s question, the response will have the flagged property set to true. In that case we will throw a general error back in the handler, but you could also inspect the response to find what categories are problematic and include more info in the error.

If the question passes moderation then we can generate the embeddings for the question.

Embeddings generation request​

For the embeddings request, we will call the openai.embeddings.create method, in a new function called getEmbedding.

// app/api/generate-answer/route.ts

async function getEmbedding(query: string) {
  const embeddingResponse = await openai.embeddings.create({
    model: "text-embedding-ada-002",
    input: query.replaceAll("\n", " "),
  });

  const [{ embedding }] = embeddingResponse.data;

  return embedding;
}

This new function again takes the question (as query). We call the OpenAI library’s embeddings.create method, specifying the model to use for generation (the model property of the options passed to the method) and passing the input (query with all newlines replaced by single spaces).

Get related documentation sections request​

Let’s look at the database query that will give us back the related sections in a variable named getSectionsQuery.

// app/api/generate-answer/route.ts

const getSectionsQuery = e.params(
    {
        target: e.OpenAIEmbedding,
        matchThreshold: e.float64,
        matchCount: e.int16,
        minContentLength: e.int16,
    },
    (params) => {
        return e.select(e.Section, (section) => {
        const dist = e.ext.pgvector.cosine_distance(
            section.embedding,
            params.target
        );
        return {
            content: true,
            tokens: true,
            dist,
            filter: e.op(
                e.op(
                  e.len(section.content),
                  ">",
                  params.minContentLength
                ),
                "and",
                e.op(dist, "<", params.matchThreshold)
            ),
            order_by: {
                expression: dist,
                empty: e.EMPTY_LAST,
            },
            limit: params.matchCount,
        };
        });
    }
);

In the above code, we use EdgeDB’s TypeScript query builder to create a query. The query takes a few parameters:

• target: Embedding array to compare against to find related sections. In this case, these will be the questions’s embeddings we just generated.
• matchThreshold: Similarity threshold. Only matches with a similarity score below this threshold will be returned. This will be a number between 0.0 and 2.0. Values closer to 0.0 mean the documentation sections must be very similar to the question while values closer to 2.0 allow for more variance.
• matchCount: Maximum number of sections to return
• minContentLength: Minimum number of characters the sections should have in order to be considered

We write a select query by calling e.select and passing it the type we want to select (e.Section). We return from that function an object representing the shape we want back plus any other clauses we need: in this case, a filter, ordering, and limit clause.

We use the cosine_distance function to calculate the similarity between the user’s question and our documentation sections. We have access to this function through EdgeDB’s pgvector extension. We then filter on that property by comparing it to the matchThreshold value we will pass when executing the query.

We want to get back the content and number of tokens for every related section that passes the filter clause (i.e., has more than minContentLength tokens and a distance from the question embedding less than our matchThreshold). We want to order results in ascending order (which is the default) by how related they are to the question (represented as dist) and to get back, at most, matchCount sections.

We’ve written the query, but it won’t help us until we execute it. We’ll do that in the getContext function.

// app/api/generate-answer/route.ts

async function getContext(embedding: number[]) {
    const sections = await getSectionsQuery.run(client, {
        target: embedding,
        matchThreshold: 0.3,
        matchCount: 8,
        minContentLength: 20,
    });

    let tokenCount = 0;
    let context = "";

    for (let i = 0; i < sections.length; i++) {
        const section = sections[i];
        const content = section.content;
        tokenCount += section.tokens;

        if (tokenCount >= 1500) {
            tokenCount -= section.tokens;
            break;
        }

        context += `${content.trim()}\n---\n`;
    }

    return context;
}

This function takes the embeddings of the question (the embedding parameter) and returns the related documentation sections.

We start by running the query and passing in some values for the parameters:

• the question embeddings that were passed to the function
• a matchThreshold value of 0.3. You can tinker with this if you don’t like the results.
• a matchCount. We’ve chosen 8 here which represents the most sections we’ll get back.
• a minContentLength of 20 characters

We then iterate through the sections that came back to prepare them to send on to the chat completions API. This involves incrementing the token count for the current section, making sure the overall token count doesn’t exceed our maximum of 1,500 for the context (to stay under the LLM’s token limit), and, if the token count isn’t exceeded, adding the trimmed content of this section to context which we will ultimately return. Since we ordered this query by dist ascending, and since lower dist values mean more similar sections, we will be sure to get the most similar sections before we hit our token limit.

With our context ready, it’s time to get our user their answer.

Chat completions request​

Before we make our completion request, we will build the full input which consists of the user’s question, the related documentation, and the system message. The system message should tell the language model what tone to use when answering question and some general instructions on what is expected from it. With that you can give it some personality that it will bake into every response. We’ll combine all of these parts in a function called createFullPrompt.

// app/api/generate-answer/route.ts

function createFullPrompt(query: string, context: string) {
    const systemMessage = `
        As an enthusiastic EdgeDB expert keen to assist,
        respond to questions referencing the given EdgeDB
        sections.

        If unable to help based on documentation, respond
        with: "Sorry, I don't know how to help with that."`;

    return stripIndents`
        ${oneLineTrim`${systemMessage}`}

        EdgeDB sections: """
        ${context}
        """

        Question: """
        ${query}
        """`;
}

This function takes the question (as query) and the related documentation (as context), combines them with a system message, and formats it all nicely for easy consumption by the chat completions API.

We’ll pass the prompt returned from that function as an argument to a new function (getOpenAiAnswer) that will get the answer from the OpenAI and return it.

// app/api/generate-answer/route.ts

async function getOpenAiAnswer(prompt: string) {
  const completion = await openai.chat.completions
    .create({
      model: "gpt-3.5-turbo",
      messages: [{ role: "user", content: prompt }],
      max_tokens: 1024,
      temperature: 0.1,
      stream: true,
    })
    .asResponse();

  return completion;
}

Let’s take a look at the options we’re sending through:

• model: The language model we want the chat completions API to use when answering the question. (You can alternatively use gpt-4 to if you have access to it.)
• messages: We send the prompt as part of the messages property. It is possible to send the system message on the first object of the array, with role: system, but since we also have the context sections as part of the input, we will just send everything with the role user.
• max_tokens: Maximum number of tokens to use for the answer.
• temperature: Number between 0 and 2. From OpenAI’s create chat completion endpoint documentation: “Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic.”
• stream: Setting this to true will have the API stream the response

The completed route​

Now, let’s take a look at the whole thing. Copy and paste this into your app/api/generate-answer/route.ts file.

// app/api/generate-answer/route.ts

import { stripIndents, oneLineTrim } from "common-tags";
import * as edgedb from "edgedb";
import e from "dbschema/edgeql-js";
import { errors } from "../../constants";
import { initOpenAIClient } from "@/utils";

export const runtime = "edge";

const openai = initOpenAIClient();

const client = edgedb.createHttpClient();

export async function POST(req: Request) {
  try {
    const { query } = await req.json();
    const sanitizedQuery = query.trim();

    const flagged = await isQueryFlagged(query);

    if (flagged) throw new Error(errors.flagged);

    const embedding = await getEmbedding(query);

    const context = await getContext(embedding);

    const prompt = createFullPrompt(sanitizedQuery, context);

    const answer = await getOpenAiAnswer(prompt);

    return new Response(answer.body, {
      headers: {
        "Content-Type": "text/event-stream",
      },
    });
  } catch (error: any) {
    console.error(error);

    const uiError = error.message || errors.default;

    return new Response(uiError, {
      status: 500,
      headers: { "Content-Type": "application/json" },
    });
  }
}

async function isQueryFlagged(query: string) {
  const moderation = await openai.moderations.create({
    input: query,
  });

  const [{ flagged }] = moderation.results;

  return flagged;
}

async function getEmbedding(query: string) {
  const embeddingResponse = await openai.embeddings.create({
    model: "text-embedding-ada-002",
    input: query.replaceAll("\n", " "),
  });

  const [{ embedding }] = embeddingResponse.data;

  return embedding;
}

const getSectionsQuery = e.params(
  {
    target: e.OpenAIEmbedding,
    matchThreshold: e.float64,
    matchCount: e.int16,
    minContentLength: e.int16,
  },
  (params) => {
    return e.select(e.Section, (section) => {
      const dist = e.ext.pgvector.cosine_distance(
        section.embedding,
        params.target
      );
      return {
        content: true,
        tokens: true,
        dist,
        filter: e.op(
          e.op(
            e.len(section.content),
            ">",
            params.minContentLength
          ),
          "and",
          e.op(dist, "<", params.matchThreshold)
        ),
        order_by: {
          expression: dist,
          empty: e.EMPTY_LAST,
        },
        limit: params.matchCount,
      };
    });
  }
);

async function getContext(embedding: number[]) {
  const sections = await getSectionsQuery.run(client, {
    target: embedding,
    matchThreshold: 0.3,
    matchCount: 8,
    minContentLength: 20,
  });

  let tokenCount = 0;
  let context = "";

  for (let i = 0; i < sections.length; i++) {
    const section = sections[i];
    const content = section.content;
    tokenCount += section.tokens;

    if (tokenCount >= 1500) {
      tokenCount -= section.tokens;
      break;
    }

    context += `${content.trim()}\n---\n`;
  }

  return context;
}

function createFullPrompt(query: string, context: string) {
  const systemMessage = `
        As an enthusiastic EdgeDB expert keen to assist,
        respond to questions referencing the given EdgeDB
        sections.

        If unable to help based on documentation, respond
        with: "Sorry, I don't know how to help with that."`;

  return stripIndents`
        ${oneLineTrim`${systemMessage}`}

        EdgeDB sections: """
        ${context}
        """

        Question: """
        ${query}
        """`;
}

async function getOpenAiAnswer(prompt: string) {
  const completion = await openai.chat.completions
    .create({
      model: "gpt-3.5-turbo",
      messages: [{ role: "user", content: prompt }],
      max_tokens: 1024,
      temperature: 0.1,
      stream: true,
    })
    .asResponse();

  return completion;
}

With the route complete, we can build the UI and connect everything together.

Building the UI​

To make things as simple as possible, we will just update the Home component that’s inside app/page.tsx file. By default all components inside the App Router are server components, but we want to have client-side interactivity and dynamic updates. In order to do that we have to use a client component for our Home component. The way to accomplish that is to convert the page.tsx file to use the client component. We do that by adding the use client directive to the top of the file.

Follow along for understanding and copy/paste the full component code at the end of the section.

// app/page.tsx

"use client";

Now we build a simple UI for the chatbot.

// app/page.tsx

import { useState } from "react";
import { errors } from "./constants";

export default function Home() {
    const [prompt, setPrompt] = useState("");
    const [question, setQuestion] = useState("");
    const [answer, setAnswer] = useState<string>("");
    const [isLoading, setIsLoading] = useState(false);
    const [error, setError] = useState<string | undefined>(undefined);

    const handleSubmit = () => {};

    return (
    <main className="w-screen h-screen flex items-center justify-center bg-[#2e2e2e]">
        <form className="bg-[#2e2e2e] w-[540px] relative">
        <input
            className={`py-5 pl-6 pr-[40px] rounded-md bg-[#1f1f1f] w-full
            outline-[#1f1f1f] focus:outline outline-offset-2 text-[#b3b3b3]
            mb-8 placeholder-[#4d4d4d]`}
            placeholder="Ask a question..."
            value={prompt}
            onChange={(e) => {
              setPrompt(e.target.value);
            }}
        ></input>
        <button
            onClick={handleSubmit}
            className="absolute top-[25px] right-4"
            disabled={!prompt}
        >
            <ReturnIcon
            className={`${!prompt ? "fill-[#4d4d4d]" : "fill-[#1b9873]"}`}
            />
        </button>
        <div className="h-96 px-6">
            {question && (
            <p className="text-[#b3b3b3] pb-4 mb-8 border-b border-[#525252] ">
                {question}
            </p>
            )}
            {(isLoading && <LoadingDots />) ||
            (error && <p className="text-[#b3b3b3]">{error}</p>) ||
            (answer && <p className="text-[#b3b3b3]">{answer}</p>)}
        </div>
        </form>
    </main>
    );
}

function ReturnIcon({ className }: { className?: string }) {
    return (
        <svg
            width="20"
            height="12"
            viewBox="0 0 20 12"
            fill="none"
            xmlns="http://www.w3.org/2000/svg"
            className={className}
        >
            <path
            fillRule="evenodd"
            clipRule="evenodd"
            d={`M12 0C11.4477 0 11 0.447715 11 1C11 1.55228 11.4477 2 12
            2H17C17.5523 2 18 2.44771 18 3V6C18 6.55229 17.5523 7 17
            7H3.41436L4.70726 5.70711C5.09778 5.31658 5.09778 4.68342 4.70726
            4.29289C4.31673 3.90237 3.68357 3.90237 3.29304 4.29289L0.306297
            7.27964L0.292893 7.2928C0.18663 7.39906 0.109281 7.52329 0.0608469
            7.65571C0.0214847 7.76305 0 7.87902 0 8C0 8.23166 0.078771 8.44492
            0.210989 8.61445C0.23874 8.65004 0.268845 8.68369 0.30107
            8.71519L3.29289 11.707C3.68342 12.0975 4.31658 12.0975 4.70711
            11.707C5.09763 11.3165 5.09763 10.6833 4.70711 10.2928L3.41431
            9H17C18.6568 9 20 7.65685 20 6V3C20 1.34315 18.6568 0 17 0H12Z`}
            />
        </svg>
    );
}

function LoadingDots() {
    return (
        <div className="grid gap-2">
            <div className="flex items-center space-x-2 animate-pulse">
            <div className="w-1 h-1 bg-[#b3b3b3] rounded-full"></div>
            <div className="w-1 h-1 bg-[#b3b3b3] rounded-full"></div>
            <div className="w-1 h-1 bg-[#b3b3b3] rounded-full"></div>
            </div>
        </div>
    );
}

We have created an input field where the user can enter a question. The text the user types in the input field is captured as prompt. question is the submitted prompt that we show under the input when user submits their question. We clear the input and delete the prompt when user submits it, but keep the question value so the user can reference it.

Let’s look at the fleshed-out form submission handler function that we stubbed in earlier:

// app/page.tsx

const handleSubmit = (
  e: KeyboardEvent | React.MouseEvent<HTMLButtonElement>
) => {
  e.preventDefault();

  setIsLoading(true);
  setQuestion(prompt);
  setAnswer("");
  setPrompt("");
  generateAnswer(prompt);
};

When the user submits a question, we set the isLoading state to true and show the loading indicator. We clear the prompt state and set the question state. We also clear the answer state because the answer may hold an answer to a previous question, but we want to start with an empty answer.

At this point we want to create a server-sent event and send a request to our api/generate-answer route. We will do this inside the generateAnswer function.

The browser-native SSE API doesn’t allow the client to send a payload to the server; the client is only able to open a connection to the server to begin receiving events from it via a GET request. In order for the client to be able to send a payload via a POST request to open the SSE connection, we will use the sse.js package, so let’s install it.

npm install sse.js

This package doesn’t have a corresponding types package, so we need to add them manually. Let’s create a new folder named types in the project root and an sse.d.ts file inside it.

mkdir types && touch types/sse.d.ts

Open sse.d.ts and add this code:

// types/sse.d.ts

type SSEOptions = EventSourceInit & {
    payload?: string;
};

declare module "sse.js" {
    class SSE extends EventSource {
        constructor(url: string | URL, sseOptions?: SSEOptions);
        stream(): void;
    }
}

This extends the native EventStream by adding a payload to the constructor. We also added the stream function to it which is used to activate the stream in the sse.js library.

This allows us to import SSE in page.tsx and use it to open a connection to our handler route while also sending the user’s query.

// app/page.tsx
"use client";

import { useState } from "react";
import { useState, useRef } from "react";
import { SSE } from "sse.js";
import { errors } from "./constants";

export default function Home() {
    const eventSourceRef = useRef<SSE>();

    const [prompt, setPrompt] = useState("");
    const [question, setQuestion] = useState("");
    const [answer, setAnswer] = useState<string>("");
    const [isLoading, setIsLoading] = useState(false);
    const [error, setError] = useState<string | undefined>(undefined);

    const handleSubmit = () => {};

    const generateAnswer = async (query: string) => {
        if (eventSourceRef.current) eventSourceRef.current.close();

        const eventSource = new SSE(`api/generate-answer`, {
            payload: JSON.stringify({ query }),
        });
        eventSourceRef.current = eventSource;

        eventSource.onerror = handleError;
        eventSource.onmessage = handleMessage;
        eventSource.stream();
    };

    handleError() { /* … */ }
    handleMessage() { /* … */ }
// …

Note that we save a reference to the eventSource object. We need this in case a user submits a new question while answer to the previous one is still assembling on the client. If we don’t close the existing connection to the server before opening the new one, this could cause problems since two connections will be open and trying to receive data.

We opened a connection to the server, and we are now ready to receive events from it. We just need to write handlers for those events so the UI knows what to do with them. We will get the answer as part of a message event, and if an error is returned, the server will send an error event to the client.

Let’s break down these handlers.

// app/page.tsx

// …

function handleError(err: any) {
    setIsLoading(false);

    const errMessage =
    err.data === errors.flagged ? errors.flagged : errors.default;

    setError(errMessage);
}

function handleMessage(e: MessageEvent<any>) {
    try {
        setIsLoading(false);
        if (e.data === "[DONE]") return;

        const chunkResponse = JSON.parse(e.data);
        const chunk = chunkResponse.choices[0].delta?.content || "";
        setAnswer((answer) => answer + chunk);
    } catch (err) {
        handleError(err);
    }
}

When we get the message event, we extract the data from it and add it to the answer state until we receive all chunks. This is indicated when the data is equal to [DONE], meaning the whole answer has been received and the connection to the server will be closed. There is no data to be parsed in this case, so we return instead of trying to parse it. (An error will be thrown if we try to parse it in this case.)

The completed UI​

Put all that together, and you have this (which can be copy/pasted to app/page.tsx):

// app/page.tsx

"use client";

import { useState, useRef } from "react";
import { SSE } from "sse.js";
import { errors } from "./constants";

export default function Home() {
  const eventSourceRef = useRef<SSE>();

  const [prompt, setPrompt] = useState("");
  const [question, setQuestion] = useState("");
  const [answer, setAnswer] = useState<string>("");
  const [isLoading, setIsLoading] = useState(false);
  const [error, setError] = useState<string | undefined>(undefined);

  const handleSubmit = (
    e: KeyboardEvent | React.MouseEvent<HTMLButtonElement>
  ) => {
    e.preventDefault();

    setIsLoading(true);
    setQuestion(prompt);
    setAnswer("");
    setPrompt("");
    generateAnswer(prompt);
  };

  const generateAnswer = async (query: string) => {
    if (eventSourceRef.current) eventSourceRef.current.close();

    const eventSource = new SSE(`api/generate-answer`, {
      payload: JSON.stringify({ query }),
    });
    eventSourceRef.current = eventSource;

    eventSource.onerror = handleError;
    eventSource.onmessage = handleMessage;
    eventSource.stream();
  };

  function handleError(err: any) {
    setIsLoading(false);

    const errMessage =
      err.data === errors.flagged ? errors.flagged : errors.default;

    setError(errMessage);
  }

  function handleMessage(e: MessageEvent<any>) {
    try {
      setIsLoading(false);
      if (e.data === "[DONE]") return;

      const chunkResponse = JSON.parse(e.data);
      const chunk = chunkResponse.choices[0].delta?.content || "";
      setAnswer((answer) => answer + chunk);
    } catch (err) {
      handleError(err);
    }
  }

  return (
    <main className="w-screen h-screen flex items-center justify-center bg-[#2e2e2e]">
      <form className="bg-[#2e2e2e] w-[540px] relative">
        <input
          className={`py-5 pl-6 pr-[40px] rounded-md bg-[#1f1f1f] w-full
            outline-[#1f1f1f] focus:outline outline-offset-2 text-[#b3b3b3]
            mb-8 placeholder-[#4d4d4d]`}
          placeholder="Ask a question..."
          value={prompt}
          onChange={(e) => {
            setPrompt(e.target.value);
          }}
        ></input>
        <button
          onClick={handleSubmit}
          className="absolute top-[25px] right-4"
          disabled={!prompt}
        >
          <ReturnIcon
            className={`${!prompt ? "fill-[#4d4d4d]" : "fill-[#1b9873]"}`}
          />
        </button>
        <div className="h-96 px-6">
          {question && (
            <p className="text-[#b3b3b3] pb-4 mb-8 border-b border-[#525252] ">
              {question}
            </p>
          )}
          {(isLoading && <LoadingDots />) ||
            (error && <p className="text-[#b3b3b3]">{error}</p>) ||
            (answer && <p className="text-[#b3b3b3]">{answer}</p>)}
        </div>
      </form>
    </main>
  );
}

function ReturnIcon({ className }: { className?: string }) {
  return (
    <svg
      width="20"
      height="12"
      viewBox="0 0 20 12"
      fill="none"
      xmlns="http://www.w3.org/2000/svg"
      className={className}
    >
      <path
        fillRule="evenodd"
        clipRule="evenodd"
        d={`M12 0C11.4477 0 11 0.447715 11 1C11 1.55228 11.4477 2 12
            2H17C17.5523 2 18 2.44771 18 3V6C18 6.55229 17.5523 7 17
            7H3.41436L4.70726 5.70711C5.09778 5.31658 5.09778 4.68342 4.70726
            4.29289C4.31673 3.90237 3.68357 3.90237 3.29304 4.29289L0.306297
            7.27964L0.292893 7.2928C0.18663 7.39906 0.109281 7.52329 0.0608469
            7.65571C0.0214847 7.76305 0 7.87902 0 8C0 8.23166 0.078771 8.44492
            0.210989 8.61445C0.23874 8.65004 0.268845 8.68369 0.30107
            8.71519L3.29289 11.707C3.68342 12.0975 4.31658 12.0975 4.70711
            11.707C5.09763 11.3165 5.09763 10.6833 4.70711 10.2928L3.41431
            9H17C18.6568 9 20 7.65685 20 6V3C20 1.34315 18.6568 0 17 0H12Z`}
      />
    </svg>
  );
}

function LoadingDots() {
  return (
    <div className="grid gap-2">
      <div className="flex items-center space-x-2 animate-pulse">
        <div className="w-1 h-1 bg-[#b3b3b3] rounded-full"></div>
        <div className="w-1 h-1 bg-[#b3b3b3] rounded-full"></div>
        <div className="w-1 h-1 bg-[#b3b3b3] rounded-full"></div>
      </div>
    </div>
  );
}

With that, the UI can now get answers from the Next.js route. The build is complete, and it’s time to try it out!

Testing it out​

You should now be able to run the project to test it.

npm run dev

If you used our example documentation, the chatbot will know a few things about EdgeQL along with whatever it was trained on.

Some questions you might try:

• “What is EdgeQL?”
• “Who is EdgeQL for?”
• “How should I get started with EdgeQL?”

If you don’t like the responses you’re getting, here are a few things you might try tweaking:

• systemMessage in the createFullPrompt function in app/api/generate-answer/route.ts
• temperature in the getOpenAiAnswer in app/api/generate-answer/route.ts
• the matchThreshold value passed to the query from the getContext function in app/api/generate-answer/route.ts

You can see the finished source code for this build in our examples repo on GitHub. You might also find our actual implementation interesting. You’ll find it in our website repo. Pay close attention to the contents of buildTools/gpt, where the embedding generation happens and components/gpt, which contains most of the UI for our chatbot.

If you have trouble with the build or just want to hang out with other EdgeDB users, please join our awesome community on Discord!