Join Discord 👾Design DocsTrust Center

Amazon Lex V2

Suggest Edits

Overview

In this guide, we will take a look at how to setup a custom NLU connection between Lex V2 and Voiceflow.

📘

Resources

Prerequisites

Here are the tools you will need for this project:

  1. Amazon Lex V2 Bot
  2. Voiceflow Account

Setting up the Project

Install and run the project:

  1. Clone this repo:
git clone https://github.com/zslamkov/voiceflow_lexv2.git
  1. Install dependencies:
npm install

Signing AWS HTTP Requests

To allow for all of our AWS requests to be signed with an AWSv4 signature, we will be using the aws-axios package.

We can complete the connection with only a few lines of code

const client = axios.create();

const interceptor = aws4Interceptor(
  {
    region: "us-east-1",
    service: "lex",
  },
  {
    accessKeyId: process.env.AWS_KEY_ID,
    secretAccessKey: process.env.AWS_SECRET,
  }
);

client.interceptors.request.use(interceptor);

Retrieve Intent Data

We will be sending an unstructured utterance to the Lex V2 RecognizeText endpoint in order to receive a response with:

  1. The triggered intent name
  2. Entity key/values
  3. Confidence score

🚧

INTENT AND SLOT NAMES MUST MATCH

Make sure to have all slot and intent names in your Lex bot and Voiceflow interaction model match in order to trigger the right path.

The below function illustrates the request and how to transform the response into a format ready for Voiceflow. 

async function detectIntent(textInput, sessionID) {
  const response = await client
    .post(
      `https://runtime-v2-lex.us-east-1.amazonaws.com/bots/${botID}/botAliases/${aliasID}/botLocales/${botLocale}/sessions/${sessionID}/text`,
      { text: textInput }
    )
    .then(
      (response) => {
        console.log(response.data.interpretations[0].nluConfidence.score);
        return response;
      },
      (err) => {
        console.log(err);
      }
    );

  const intent = response.data.interpretations[0].intent.name;
  const entities = response.data.interpretations[0].intent.slots;
  const nluConfidence = response.data.interpretations[0].nluConfidence.score;

  const createIntent = (name, value) => ({ name, value });
  const arr = [];

  for (const [key, value] of Object.entries(entities)) {
    if (value != null) {
      const intents = createIntent(key, value.value.interpretedValue);
      arr.push(intents);
    }
  }

  return { response: intent, entities: arr, confidence: nluConfidence };
}

Node.js App

The below app allows us to interact with our chat assistant via the CLI.

We will be retrieving the userID by having the user enter their name. Once completed, we will send a launch request to Voiceflow to start the conversation and send the first steps to the channel. 

async function main() {
  const userID = await cli.prompt("> What is your name?");
  // send a simple launch request starting the dialog
  let isRunning = await interact(userID, { type: "launch" });

Now on each new interaction from the user, we will be passing the userID and text input through the detectIntent function and returning the intent object that will be used in the next line of code to populate the request payload to Voiceflow. 

  while (isRunning) {
    const nextInput = await cli.prompt("> Say something");
    // send a simple text type request with the user input
    let intent = await detectIntent(nextInput, userID);
    
    isRunning = await interact(userID, {
      type: "intent",
      payload: {
        intent: {
          name: intent.response,
        },
        entities: intent.entities,
        confidence: intent.confidence,
      },
    });
  }
  console.log("The end! Start me again with `npm start`");
}

main();

Run

What it might look like in action:

$ npm start

> What is your name?: zoran
what can I do for you?
...
> Say something: send email
who is the recipient?
...
> Say something: [email protected]
what is the title of your email?
...
> Say something: How was your day?
sending the email for [email protected] called "How was your day?". Is that correct?
...
> Say something: yes
successfully sent the email for [email protected] called "How was your day?"
The end! Start me again with `npm start`

Updated 5 months ago