Imagine you could build a video AI assistant that humans can interact with in real time. Now, imagine you could do it in five minutes. Well, now you can. Build a real time conversations with AI digital twins with just an API call.
In this tutorial, we’ll cover three of the main pieces that can get you off to the races:
We cover:
- Code Snippet to Build Conversational AI Overview
- How to Choose a Conversational Replica
- Give Your AI Context
- Properties for Controlling the Conversation
- Summary of Building Your First Real Time Conversational AI
Code Snippet to Build Conversational AI Overview
Before we dive into the best practices, here’s what the code looks like. Before you work through this example, you need to pip install requests python-dotenv and sign up for a Tavus Account. Once you've signed up for an account, you can navigate to the key icon on the left side and then click "Create New Key" to get your new API key. Make sure you copy and save the key once you get it!
import requests
url = "https://tavusapi.com/v2/conversations"
The URL we’ll hit for this API is shown above, and an example payload is shown below. We’ll cover how to use each of the pieces of the API call in this example with a detailed walkthrough of the different parameters in the sections below.
payload = {
"replica_id": "<string>",
"conversation_name": "<string>",
"conversational_context": "<string>",
"properties": {
"max_call_duration": 240,
"participant_left_timeout": 0,
"enable_recording": True,
"recording_s3_bucket_name": "<string>",
"recording_s3_bucket_region": "<string>",
"aws_assumed_role_arn": "<string>"
}
}
Before we make our API call, we need to load our API key. I use dotenv to handle environment variables. Once we have our API key, we put it in the header and set our Content-Type to “application/json”. From there, we are ready to make a POST request to the URL specified above with the payload and headers we’ve created in this code snippet.
from dotenv import load_dotenv
import os
load_dotenv()
TAVUS_API_KEY = os.environ["TAVUS_API_KEY"]
headers = {
"x-api-key": TAVUS_API_KEY,
"Content-Type": "application/json"
}
response = requests.request("POST", url, json=payload, headers=headers)
print(response.text)
When we get a response, it will look something like this:
{
"conversation_id": "c2b10f44",
"conversation_name": "New Conversation 1722996608090",
"conversation_url": "https://tavus.daily.co/c2b10f44",
"status": "active",
"callback_url": null,
"created_at": "2024-08-07T02:10:08.103Z"
}
The main thing to pay attention to here is “conversation_url”. You can click this URL and it will take you directly to a meeting room where you or a user can chat with the digital replica in real time.
Choose a Conversational Replica
The first thing to do when creating a conversation is to choose a conversational replica. We can specify this in the first line in the payload under replica_id.
There are two options here:
- You can create your own conversational replica (aka your digital twin) with an API call
- You can choose a stock replica
If you want to create your own, you can do so with an API call and a simple video.
After you pick a conversational replica, the second line is conversation_name. This is where you can name your conversation.
Give Your AI Context
The next line we see is conversational_context. Conversational context is the last mile situational context you need to give to the AI so it knows how to conduct the conversation. For example, if you’re building a sales coach, which we will cover an example of in our next blog, you would want to give the AI context on the the sales meeting.
Examples of conversational context could be:
- This person set a meeting with you to talk about video AI. Ask them about the value propositions and get them to ask you about your problem set.
- This is a conversation with a friend. Your friend is seeking some advice on how to approach looking for a job.
- You have set a meeting with this person to discuss a potential real estate investment. Here are the details of the deal: <x y z details>. See if this is a fit for your prospect.
Properties for Controlling the Conversation
The final block of the payload we need to understand is the properties block. There are six different properties available for you to control. Let’s take a look at what each of them does.
- "max_call_duration" controls how long the maximum length of the call can be *in seconds*, so 3600 as shown in the example is an hour long call
- "participant_left_timeout" controls how long the call will last after a participant leaves *in seconds* so 60 shown is a one minute timeout
- "enable_recording" decides whether or not it’s possible for you to record the call. The next three properties are only used when this property is set to true
- "recording_s3_bucket_name" is the name of the S3 bucket where you want to save your recording
- "recording_s3_bucket_region" is the region for the S3 bucket where you want to save your recording
- "aws_assumed_role_arn" is the role you need to assume to save your recording
Summary of Building Your First Real Time AI Video Conversation
In this article we learned how you can quickly get started with building a real time conversational AI. All you need is an internet connection, some Python knowledge, and an API key from Tavus. With these three things in hand, you can simply set off a POST request with a series of parameters, and get a link to a conversation back.
We also covered what many of the important parameters are. Namely, we covered:
- The
replica_id, which allows you to pick a replica to use in a conversation - The
conversational_context, which sets the specific context for a conversation - The six
propertiesof a conversation you can control
Next time, we’ll dive further into building a custom conversational AI by showing you how you can create a "persona", which can provide more than just context for your conversation, but also a background and set of expertise for the replica, for example a sales agent.
.png)