WEBVTT 00:00.160 --> 00:00.960 Welcome! 00:00.960 --> 00:05.280 In this tutorial, we are going to build a personal assistant that can make phone calls for you. 00:05.560 --> 00:06.840 So let's get started. 00:06.840 --> 00:13.240 You have already seen this AI voice assistant built in and integrated into an annotation workflow in 00:13.240 --> 00:18.000 one of the demos, but this time we are going to do something slightly different. 00:18.600 --> 00:20.880 First we are going to watch the full demo. 00:20.920 --> 00:26.520 Then I'm going to give you an overview of the setup and explain how it works, note by note. 00:26.520 --> 00:31.400 And instead of building it completely from scratch, we are going to use the template provided in the 00:31.400 --> 00:36.600 resources of this lesson and I will show you exactly what you need to change to make it work for you. 00:37.360 --> 00:43.000 This way you will learn how to take advantage of the templates we provide, and also how to adapt blueprints 00:43.120 --> 00:44.320 to fit your own needs. 00:44.760 --> 00:49.680 Because personally, I believe that the best way to learn any ten after you've got the basics is by 00:49.680 --> 00:55.880 modifying, recreating, and rebuilding templates so that the supervisor agent from your workflow can 00:55.880 --> 00:57.840 connect to it and take action. 00:57.840 --> 00:59.320 So now let's watch the demo. 00:59.560 --> 01:00.280 Hi everyone. 01:00.410 --> 01:04.330 Now I'm going to show you my personal assistant agent in action. 01:06.530 --> 01:07.050 Hi, Emily. 01:07.050 --> 01:12.890 Can you please call Mark and ask if we can meet tomorrow at 7 a.m. in the office to discuss the onboarding 01:12.890 --> 01:14.130 of two new clients. 01:14.570 --> 01:18.850 And if that works for him, send him a confirmation email and also update my calendar. 01:21.890 --> 01:24.850 So I should receive a call in a second from my personal assistant. 01:29.770 --> 01:36.570 So the main agent is looking for Mark's mobile number in the contacts data stored in a Google sheet, 01:37.370 --> 01:39.170 and is calling a subagents. 01:43.410 --> 01:43.810 Yep. 01:48.290 --> 01:50.850 Hi, I'm Emily, Damien's personal assistant. 01:51.370 --> 01:51.930 Hi, Emily. 01:51.930 --> 01:52.930 How can I help you? 01:54.370 --> 01:54.890 Hi, Mark. 01:54.890 --> 02:00.180 Damien asks if you can meet tomorrow at 7 a.m. in the office to discuss the onboarding of two new clients. 02:00.860 --> 02:01.180 Yes. 02:01.180 --> 02:01.740 No problem. 02:01.780 --> 02:03.140 That works for me. 02:03.180 --> 02:03.620 Great. 02:04.700 --> 02:05.260 Great. 02:05.500 --> 02:06.540 I'll inform Damien. 02:07.100 --> 02:07.620 Thank you. 02:07.660 --> 02:11.100 Can you ask him if I need to prepare anything? 02:12.820 --> 02:15.220 I will check with Damien and get back to you shortly. 02:15.780 --> 02:16.260 Thank you. 02:16.260 --> 02:16.780 Bye bye. 02:19.500 --> 02:19.900 Yes. 02:19.900 --> 02:24.060 So now the supervisor agent is calling the Subagents. 02:24.060 --> 02:25.780 So email and calendar agents. 02:27.900 --> 02:31.700 And I should get a confirmation email and have a meeting booked in my calendar. 02:33.540 --> 02:35.020 So let's check my inbox. 02:38.340 --> 02:38.580 Yeah. 02:38.580 --> 02:39.300 That's great. 02:39.300 --> 02:40.300 And my calendar. 02:41.940 --> 02:42.220 Yeah. 02:42.220 --> 02:42.900 That's awesome. 02:43.740 --> 02:44.300 All right. 02:44.300 --> 02:48.700 So now I'm going to walk you through each node and explain everything. 02:49.620 --> 02:55.020 I communicate with Emily via telegram because it's the easiest way and it's also free. 02:55.300 --> 03:01.910 But you can easily integrate with other apps such as WhatsApp, slack, etc. as you can see here in 03:01.910 --> 03:06.670 telegram, I got a confirmation from my agent summarizing everything. 03:07.110 --> 03:15.350 She also passed on what I asked her as Mark confirmed that she sent an email to Mark, updated my calendar, 03:15.350 --> 03:18.070 and even added a link to make it easy to access. 03:18.110 --> 03:25.670 Apart from the three subagents so phone call, calendar and email agents, I've added a vector database 03:25.670 --> 03:27.550 as another tool for Emily. 03:27.990 --> 03:32.070 So in short, Emily is trained on Google Docs stored in my Google Drive. 03:34.190 --> 03:38.710 So this is my company's knowledge base. 03:39.070 --> 03:43.390 So if I ask her, for example, what is adaptive AI? 03:48.670 --> 03:54.150 She knows exactly what's included in that document and can answer straight away. 03:55.070 --> 04:01.520 This means she can remind me about my agency SOPs past project details, training materials and much 04:01.520 --> 04:01.840 more. 04:02.560 --> 04:03.720 So she is very helpful. 04:04.240 --> 04:11.520 You can add more tools to Emily like want to do research or analyze competitors or create reports. 04:11.560 --> 04:15.800 The very first step is this trigger node which listens for incoming events. 04:16.440 --> 04:22.400 So every time I send a voice or text message to Emily, the workflow is activated thanks to that node. 04:22.520 --> 04:29.160 So when a new event is detected, the trigger captures it and pass it to it along to the rest of the 04:29.160 --> 04:29.800 workflow. 04:30.680 --> 04:34.440 So any interaction with Emily on Telegram automatically starts the process. 04:34.800 --> 04:35.120 Great. 04:35.160 --> 04:39.320 Now, with this in place, Emily is officially ready to listen to us on telegram. 04:40.840 --> 04:49.840 So next, once our telegram trigger has captured an event, this node determines content type. 04:50.200 --> 04:58.650 So acts as a decision maker and analyzes the incoming messages to figure out whether it's a text, voice 04:58.650 --> 05:00.290 note or something else. 05:00.810 --> 05:08.170 So this node uses simple rules to check for key properties in the coming data. 05:08.610 --> 05:15.530 For example, if the message contains a text field, we know it's a regular text message, and if it 05:15.530 --> 05:21.930 has a file ID in the voice field, it's an audio message and so on. 05:24.250 --> 05:27.730 So based on these rules, the workflow branches out. 05:28.330 --> 05:30.730 Text messages go one way. 05:33.010 --> 05:33.970 Audio another. 05:34.890 --> 05:39.250 And errors like unsupported content types take a separate path. 05:40.970 --> 05:43.410 This ensures Emily can handle every input. 05:43.890 --> 05:51.530 For example, if it's a voice message, the workflow prepares to download and transcribe it. 05:52.850 --> 05:56.930 If it's text, it moves forward to process the message. 05:56.930 --> 06:00.660 Search, so every type of input is handled in the right way. 06:01.220 --> 06:13.460 Now when someone sends a voice message download voice file node, takes the file ID from the incoming 06:13.620 --> 06:19.740 audio message and known as the actual audio directly from the telegram. 06:22.660 --> 06:23.060 Great. 06:23.060 --> 06:29.460 So now when the voice message is ready for the next phase of the workflow, the file is prepared so 06:29.460 --> 06:36.900 it can be transcribed into text and the AI agent can understand its content. 06:37.420 --> 06:45.820 This node uses OpenAI's whisper model to to do the transcription, and this model works great even if 06:45.820 --> 06:48.420 there is an accent or some background noise. 06:48.940 --> 06:55.220 So the node takes the audio file, processes it, and spits out plain text. 06:57.100 --> 07:03.700 For example, if I send a voice message saying what's on my calendar tomorrow, the note turns that 07:03.700 --> 07:04.900 into a text message. 07:06.380 --> 07:11.540 This text is what Emily uses to figure out what you want her to do next. 07:13.820 --> 07:14.940 So now let's move on. 07:16.980 --> 07:18.020 The next note. 07:19.060 --> 07:25.740 Combine content so you have the input and it can be text or a transcribed voice message right. 07:26.180 --> 07:28.740 So this note pull everything together. 07:28.780 --> 07:30.020 Add some labels. 07:31.900 --> 07:37.860 So it combines it into a single variable called combined message. 07:38.980 --> 07:44.980 This way no matter how the message was sent, Emily processed it in the same way. 07:45.660 --> 07:53.340 So it figures out the type of message, like if it's a text query or a voice message, and tags it with 07:53.540 --> 07:54.380 message type. 07:55.380 --> 08:05.670 So if the message was --, it adds a little label in source type to know that now that everything's 08:05.710 --> 08:09.710 sorted, we are ready to send it over to Emily's brain. 08:10.590 --> 08:12.110 So the main agent. 08:14.790 --> 08:18.910 Now, Emily processes the message and decides what needs to be done. 08:19.470 --> 08:29.190 So basically, Emily works as a as a supervisor agent and delegates tasks to the three sub agents working 08:29.190 --> 08:29.830 under her. 08:31.430 --> 08:34.430 So first make phone call agent. 08:35.310 --> 08:41.710 If you ask Emily to make a call, she passes the details like the name, phone number and what to say 08:42.110 --> 08:44.350 to this agent to handle it. 08:45.870 --> 08:47.350 Next calendar agent. 08:47.870 --> 08:53.190 So for anything related to your schedule like checking availability, adding, or checking events, 08:53.590 --> 08:55.550 Emily hands it over to this agent. 08:56.760 --> 08:57.720 And the last one? 08:58.080 --> 08:59.000 Female agent. 08:59.600 --> 09:01.000 This one handles emails. 09:01.520 --> 09:05.360 So whether it's sending, replying or setting up drafts. 09:07.160 --> 09:16.560 But Emily is smart, so she she always checks the contacts data, a Google sheet in Google Drive to 09:16.560 --> 09:19.920 make sure the email address is valid before sending anything. 09:21.280 --> 09:29.560 Now, really important part of this agent and how it works is the prompt. 09:30.680 --> 09:36.320 So it's basically the set of instructions that tells Emily how to handle different tasks. 09:37.280 --> 09:44.000 So the prompt is super specific and even includes variables like the current date and others. 09:44.000 --> 09:47.120 So Emily always has the right context for what she is doing. 09:49.760 --> 09:51.200 So as you can see, it's very detailed. 09:53.520 --> 09:57.370 For example, it reminds her to check the contact list. 09:59.410 --> 10:06.850 Before sending an email, or to make sure phone numbers are formatted correctly before making a call, 10:07.210 --> 10:08.810 such as adding a country code. 10:09.610 --> 10:17.810 So once again, Emily works as the supervisor agent and delegates tasks to the three sub agents, and 10:19.010 --> 10:23.570 each of these sub agents is included in separate workflows. 10:24.250 --> 10:27.530 So this keeps everything modular and easy to manage. 10:27.570 --> 10:29.130 Like in object oriented programming. 10:30.650 --> 10:33.690 So let's start with a make phone call agent. 10:34.610 --> 10:40.970 This agent uses VPI, which is a voice API platform to make the calls. 10:41.570 --> 10:47.330 When Emily needs to handle a phone call request, she gathers all the necessary details like the name 10:47.330 --> 10:55.540 of the person and the type of contact, like if it's a friend or business, and specific instructions 10:55.540 --> 10:59.860 for the call, such as ask if they are available to me tomorrow. 11:02.460 --> 11:10.860 So once Emily has all this information, she passes it to the make one call agent, which then triggers 11:10.900 --> 11:13.660 a workflow called make a phone call path. 11:14.340 --> 11:16.660 So this is the make a phone call agent. 11:19.300 --> 11:23.020 This agent uses the API to handle the actual call. 11:23.580 --> 11:27.940 And after the call is made, the response from the workflow comes back to Emily. 11:28.140 --> 11:31.340 So she knows what happened and can continue with the task. 11:32.500 --> 11:41.180 So this setup ensures that each Subagent focuses on its specific job, while Emily oversees everything 11:41.180 --> 11:44.020 and makes sure it all works together as expected. 11:45.940 --> 11:48.980 So next let's look at the calendar agent. 11:50.260 --> 11:52.260 So this is our calendar agent. 11:52.940 --> 11:58.110 So when Emily handles anything related to scheduling, like checking availability or adding events, 11:58.510 --> 12:00.310 she passes the details to this agent. 12:01.270 --> 12:05.790 Basically, this separate workflow is dedicated to managing your calendar. 12:06.190 --> 12:12.390 It processes the request, updates or retrieves the event information and sends the results back to 12:12.430 --> 12:12.950 Emily. 12:14.030 --> 12:19.910 From there, Emily uses the response to provide you with the updates or confirmations. 12:20.910 --> 12:22.110 Now the email agent. 12:23.070 --> 12:30.630 So Emily first checks the contacts data, a Google sheet to make sure the recipient's email address 12:30.670 --> 12:31.350 is valid. 12:32.070 --> 12:38.710 And once verified, the email agent triggers a separate workflow to handle the email task. 12:39.230 --> 12:45.070 So the workflow sends the email and then returns a confirmation and response back to Emily. 12:45.670 --> 12:48.470 Now let's look at the knowledge base node. 12:50.630 --> 12:59.520 So this node is connected to find Convector store, which allows Emily to access and retrieve specific 12:59.520 --> 13:01.960 information stored in database. 13:03.000 --> 13:05.840 So you can store all sorts of documents in pinecone. 13:06.320 --> 13:09.480 This could be Google Docs, spreadsheets like Google Sheets. 13:09.520 --> 13:18.400 Company policies, etc. so you can give Emily access to, for example, frequently ask questions, client 13:18.440 --> 13:23.920 details, SOPs, training materials, or any other knowledge base. 13:25.000 --> 13:33.280 So Emily can retrieve this content and fetch accurate answers by pulling directly from your stored information. 13:34.600 --> 13:41.800 For example, if you ask Emily something about your company or any stored knowledge, this node searches 13:41.800 --> 13:50.360 through the vector database for the most relevant information and sends it back, and the response is 13:50.360 --> 13:51.440 then passed to Emily. 13:51.440 --> 13:54.130 And based on that, she can provide a relevant answer. 13:57.130 --> 14:00.770 So the pinecone vector star node is where all the knowledge is kept. 14:01.170 --> 14:06.410 It stores the information in a way that makes it easy to search and find later. 14:06.730 --> 14:12.410 So all the knowledge is converted into numerical representations called embeddings. 14:14.770 --> 14:16.490 So I can understand it. 14:18.210 --> 14:24.690 Embeddings are small, easy to search versions of your data that can help large language models find 14:24.690 --> 14:25.890 exactly what they need. 14:27.010 --> 14:35.690 And when Emily sends something, this node searches through the embeddings to find the best match and 14:35.690 --> 14:37.930 sends the results to the knowledge base back. 14:38.290 --> 14:45.330 So from there, it's formatted and passed to Emily so she can use it to answer a question. 14:45.610 --> 14:51.010 So this node generates embeddings for the data stored in a Python vector store. 14:52.100 --> 14:58.060 And whenever you add new information to the database, this node processes it, creating embeddings 14:58.060 --> 15:00.580 that make it easy to search and retrieve later. 15:00.900 --> 15:03.740 So it's a key part of keeping the knowledge base up to date. 15:05.980 --> 15:09.700 Now let's talk about the window buffer memory node. 15:11.740 --> 15:12.300 This one. 15:14.460 --> 15:20.820 This node helps Emily keep track of the context during conversations by storing short term memory. 15:21.500 --> 15:27.380 It allows Emily to remember recent interactions so she can handle follow up questions without losing 15:27.420 --> 15:29.340 track of what you are talking about. 15:29.500 --> 15:34.180 For example, if you ask what's what's on my calendar tomorrow? 15:34.900 --> 15:38.380 And then follow up with can you add a meeting at 3:00 pm? 15:38.780 --> 15:42.100 Emily can connect those requests and respond. 15:43.500 --> 15:47.780 However, this is not a persistent memory like the Pinecone Vector store. 15:48.780 --> 15:52.510 Once the session ends, the memory from this note is cleared. 15:52.790 --> 15:57.150 So it's designed for short term use within the same conversation. 15:58.750 --> 16:06.030 Now, this OpenAI chat model is connected directly to Emily and acts as her brain. 16:06.350 --> 16:12.830 So when Emily needs to process a request like understanding a message or deciding what to do next, 16:13.030 --> 16:20.950 this node uses OpenAI large language model to interpret the input and provide a response. 16:21.430 --> 16:27.830 For example, if you ask Emily to send an email, this node helps analyze the message and figure out 16:27.830 --> 16:30.550 the context and decide what to do next. 16:31.830 --> 16:37.870 I've got a quick five minute tutorial on my channel showing how to connect all OpenAI models to your 16:38.110 --> 16:38.710 agents. 16:38.870 --> 16:40.150 So be sure to check it out. 16:43.510 --> 16:47.710 While this OpenAI chat model works with the knowledge base. 16:49.910 --> 16:59.560 When the pinecone Vector store finds the information Emily needs, this node helps it into a clear response 16:59.560 --> 17:00.760 she can share with you. 17:01.280 --> 17:10.080 For example, if you ask about company policies, this node takes the data from pinecone, processes 17:10.080 --> 17:13.800 it, and turns it into an easy to understand answer. 17:17.280 --> 17:21.080 And in this setup, both openly chat model nodes. 17:22.080 --> 17:28.440 This one and this one use GPT for Omni. 17:28.920 --> 17:30.280 So it's very cheap to run. 17:30.840 --> 17:33.320 I managed to achieve this with a complex prompt. 17:33.920 --> 17:36.680 Now this is the response to me node. 17:37.880 --> 17:45.240 So after Emily completes a task or finds an answer, this node sends a response back to you on telegram. 17:45.720 --> 17:50.490 Now go ahead and download all the files from the resources section of this lesson. 17:50.690 --> 17:56.970 Once you've got them, import the first file called your personal assistant into a new workflow. 17:57.370 --> 18:01.530 If you miss the lesson on how to import workflows into an A-10, no worries. 18:01.530 --> 18:05.010 Just simply go to just simply create a new workflow. 18:05.050 --> 18:08.810 Go to these three dots and import from file. 18:11.410 --> 18:12.650 You should see this workflow. 18:12.650 --> 18:18.330 And now the first thing you'll need to do is connect your credentials for some of the nodes in this 18:18.330 --> 18:18.810 setup. 18:18.810 --> 18:21.890 So especially all the telegram nodes. 18:22.210 --> 18:27.010 So if you have watched the previous lessons, you should already know how to connect your credentials 18:27.010 --> 18:31.810 like telegram, OpenAI and any others using this workflow. 18:31.850 --> 18:38.770 In most cases, Naa10 should auto detect and update those for you as long as you have set them up before. 18:38.810 --> 18:45.890 But if anything's missing or not working, just go to the Setup credentials section in this course and 18:45.890 --> 18:51.500 you will find short tutorials showing exactly how to connect each specific service step by step. 18:51.660 --> 18:57.300 So the first node we need to connect is this node to listen for incoming events. 18:57.300 --> 19:01.020 So make sure this node is connected to your telegram account. 19:01.980 --> 19:08.900 Then if we have successfully connected your telegram simply double click on all of the telegram nodes. 19:09.220 --> 19:12.660 So N810 will automatically refresh the credentials for you. 19:18.660 --> 19:21.420 Make sure you connect the same account. 19:28.420 --> 19:30.180 And you have an output node. 19:30.900 --> 19:32.500 So respond message. 19:33.860 --> 19:34.180 All right. 19:34.180 --> 19:34.740 Great. 19:35.420 --> 19:38.820 Now make sure your OpenAI account is also connected. 19:41.380 --> 19:45.700 So we can use the whisper model to transcribe the recording. 19:47.580 --> 19:48.420 Yeah great. 19:49.780 --> 19:55.100 And also here is a brain for our supervisor agent. 19:57.540 --> 19:57.820 All right. 19:57.820 --> 19:58.340 Awesome. 20:00.700 --> 20:03.860 Also disciplinary note to use the vector database. 20:05.780 --> 20:06.260 Great. 20:09.460 --> 20:16.220 Now what we need to do we need to download these three sub workflows, import them separately and then 20:16.220 --> 20:19.180 connect it to these nodes. 20:19.420 --> 20:21.020 So our personal assistant. 20:21.220 --> 20:24.500 So supervisor agent will be able to use them as tools. 20:25.100 --> 20:28.100 First let's import the phone call agent workflow. 20:30.020 --> 20:30.740 All right. 20:32.180 --> 20:33.700 Now let's rename it to. 20:35.900 --> 20:36.660 Agent. 20:42.460 --> 20:43.540 And save it. 20:43.580 --> 20:45.620 Now let's go back to our main workflow. 20:45.940 --> 20:46.300 Alright. 20:46.300 --> 20:46.620 Great. 20:46.660 --> 20:52.070 Now we should be able to choose this workflow in the Execute workflow node. 20:52.950 --> 20:55.110 So is this one for Focal Agent. 20:59.110 --> 21:00.870 So workflow from list. 21:03.710 --> 21:04.670 And we call it. 21:07.870 --> 21:09.470 Agent workflow. 21:10.310 --> 21:11.110 Alright great. 21:11.670 --> 21:15.630 Now you have to do the same thing with the two sub workflows. 21:17.150 --> 21:18.830 So I'm going to create a workflow. 21:25.590 --> 21:27.110 Import the calendar agent. 21:28.830 --> 21:29.670 Do the same thing. 21:29.670 --> 21:30.630 Import from file. 21:32.950 --> 21:33.390 Calendar. 21:33.390 --> 21:34.030 Agent. 21:35.310 --> 21:36.110 Alright great. 21:36.670 --> 21:38.230 Now just double click on this node. 21:38.230 --> 21:44.990 If you have connected your open account area it will be automatically updated. 21:45.030 --> 21:46.550 The same for Google Calendar. 21:48.520 --> 21:52.560 Make sure you are connected to the correct Google Calendar account. 21:59.440 --> 22:01.880 Okay, now let's rename this workflow. 22:02.440 --> 22:05.040 Let's call it Personal 22:07.000 --> 22:08.400 Assistant. 22:16.760 --> 22:17.560 And hit save. 22:19.080 --> 22:21.000 Now let's go back to our workflow. 22:25.400 --> 22:31.280 And let's search for the workflow we just created in this execute workflow node. 22:40.160 --> 22:42.760 Is this one personal assistant calendar node tool. 22:44.360 --> 22:45.760 So create a new workflow. 22:58.530 --> 22:59.570 Import profile. 23:06.170 --> 23:14.850 Updated credentials and make sure you are connected to the Gmail account you want to use for this setup. 23:19.450 --> 23:20.370 Let's call it. 23:24.410 --> 23:25.210 Personal. 23:31.130 --> 23:32.010 Email tool. 23:34.650 --> 23:35.330 Itself. 23:36.930 --> 23:38.690 And let's go back to our workflow. 23:40.290 --> 23:41.610 And the last step workflow. 23:41.610 --> 23:42.610 So email agent. 23:47.540 --> 23:51.060 We call this workflow personal. 23:54.300 --> 23:58.740 Tool by perfect itself. 24:02.540 --> 24:06.540 Now the next step is to create a Google Sheet with contacts data. 24:06.860 --> 24:09.300 So please go create a Google Sheet. 24:10.100 --> 24:11.540 And mine looks like that. 24:13.180 --> 24:17.260 So I have a name, column email address and phone number. 24:17.420 --> 24:24.100 So our supervisor agent will be able to look through this Google sheet and and then find the phone number 24:24.100 --> 24:25.500 we want to call to. 24:28.020 --> 24:28.340 Alright. 24:28.340 --> 24:28.860 Perfect. 24:29.340 --> 24:30.820 Now this is really important. 24:30.820 --> 24:35.980 Also make sure you are connected to the correct Google sheet account. 24:36.020 --> 24:40.340 So our supervisor agent will be able to retrieve this contacts. 24:40.780 --> 24:45.660 Then also make sure you you choose the correct document from the list. 24:45.780 --> 24:47.350 So mine is contact database. 24:47.710 --> 24:49.590 This is how I name this Google sheet. 24:53.590 --> 24:53.990 All right. 24:54.030 --> 24:54.830 Now let's move on. 24:57.510 --> 25:00.510 Now let's start configuring our agent. 25:00.550 --> 25:07.670 So this node agent is really important because it's one of the, uh, is the one responsible for making 25:07.710 --> 25:08.390 phone calls. 25:08.870 --> 25:09.830 So let's open it. 25:14.110 --> 25:14.990 And let's scroll down. 25:18.590 --> 25:22.110 As you can see here, I added a little bit of code, but don't worry. 25:22.350 --> 25:24.110 This is just a simple code snippet. 25:24.470 --> 25:30.510 So this node triggers a separate workflow that's dedicated to handling the phone call using the API. 25:30.870 --> 25:35.710 And we already imported this workflow and connect it to our supervisor agent. 25:35.950 --> 25:43.870 So here we send it a JSON object with all the important details such as who to call, the phone number, 25:44.390 --> 25:48.120 what to say and how the voice should sound. 25:49.480 --> 25:52.240 So, for example, Tom here friendly as a friendly. 25:53.520 --> 26:00.240 So these are dynamic variables which will be passed to our AI voice agent on platform. 26:00.440 --> 26:06.920 So it passes this info to the phone call agent workflow, which makes the actual call through the API 26:07.200 --> 26:08.920 and sends the response back. 26:09.360 --> 26:16.560 So this node acts like a bridge between your supervisor agent and the AI voice agent on VPI. 26:16.880 --> 26:24.320 So it delegates the task and waits for the result and lets the rest of the automation continue based 26:24.320 --> 26:26.280 on what happened in the call. 26:26.560 --> 26:26.920 Perfect. 26:26.920 --> 26:28.280 Let's close it for now. 26:30.320 --> 26:31.160 Oh, actually. 26:34.080 --> 26:38.440 You can change the callback response because for now it's. 26:39.440 --> 26:42.400 I'm not sure, but I'll check with Damien and get back to you. 26:43.600 --> 26:45.970 So in case our agent is not. 26:50.370 --> 26:57.570 Sure how to respond to a specific question that in the call it will say this so you can change it to 26:57.610 --> 26:58.090 your name. 26:58.130 --> 26:58.450 All right. 26:58.490 --> 26:58.970 Awesome. 26:59.690 --> 27:00.130 All right. 27:00.170 --> 27:06.210 Now let's jump into VPI and create our AI voice agent, the one that will actually make phone calls 27:06.210 --> 27:06.770 for us. 27:06.770 --> 27:13.850 So we are going to set up a new advanced agent, which means it will be triggered by our workflow to 27:13.890 --> 27:21.690 call someone exactly by this workflow, say what we tell it to say and send the response back. 27:21.690 --> 27:27.890 Now, in order to create your account on VPI, please use the link included in the PDF called Essential 27:27.930 --> 27:29.050 Tools for this course. 27:29.370 --> 27:32.730 You will find it in the resources section of this lesson. 27:33.290 --> 27:40.330 So by signing up through that link, you will get 1003 minutes to use with your agents on VPI, which 27:40.330 --> 27:46.340 is more than enough to build, test and run a lot of AI voice agents Once you have signed up and logged 27:46.340 --> 27:51.500 in, you will be ready to create your agent and connect it to your workflow. 27:51.500 --> 27:52.180 So let's move on. 27:52.220 --> 27:56.700 Once you create your account on VPI and then log in, you will see this dashboard. 27:56.700 --> 27:59.540 So now on the left sidebar click on assistance. 27:59.540 --> 28:03.020 So here you will build your first agent. 28:07.340 --> 28:08.700 Now let's create assistant. 28:09.780 --> 28:10.780 So hit that button. 28:12.900 --> 28:13.820 And give it a name. 28:14.540 --> 28:17.420 For example agent a voice agent. 28:20.180 --> 28:20.820 Anything. 28:25.700 --> 28:30.460 And eight n and we want to start from a blank template. 28:31.860 --> 28:33.100 So create assistant. 28:36.580 --> 28:36.940 Alright. 28:36.940 --> 28:37.380 Great. 28:37.540 --> 28:38.900 You will land on this dashboard. 28:38.900 --> 28:41.580 And let me quickly walk you through what you are seeing. 28:42.060 --> 28:50.390 So at the top you will see the cost per minute and the latency which shows how fast your agent responds. 28:51.190 --> 28:54.350 So for most use cases, these defaults are totally fine. 28:54.750 --> 29:01.710 And below that, under the model section you will choose the brain of your assistant. 29:01.830 --> 29:07.670 So here we are using OpenAI as the provider and GPT four as the model, which is very powerful. 29:07.670 --> 29:13.830 And I recommend to use this model for the best results, but you can easily switch to other, more affordable 29:13.830 --> 29:14.430 models. 29:14.830 --> 29:18.430 For now, let's stick with GPT four for now. 29:19.150 --> 29:24.630 The first message is what your agent says when a call starts. 29:25.190 --> 29:31.830 So, as you can see, the first message that the assistant will say, this can also be a URL to containerize 29:31.830 --> 29:32.630 audio file. 29:34.870 --> 29:37.990 So something like hi. 29:41.590 --> 29:42.350 I'm Emily. 29:45.720 --> 29:46.800 Personal assistant. 29:50.280 --> 29:53.160 And the most important part is the system prompt. 29:53.400 --> 30:00.040 This is where you tell your agent how to act, what tone to use, and how to handle different situations. 30:00.240 --> 30:06.840 And we also will be using um, dynamic variables we defined in the workflow. 30:07.520 --> 30:13.600 Now what I'm going to do I'm going to use the prompt I've added to the resources section of this lesson. 30:14.560 --> 30:21.800 So the prompt and copy it to the agent system prompt, we just we just created on the platform. 30:23.320 --> 30:23.800 All right. 30:23.800 --> 30:29.000 So as a role we have your Emily a voice personal assistant for Damien. 30:29.880 --> 30:34.520 And now remember the code snippet we have in our phone call agent sub workflow. 30:35.320 --> 30:35.720 So. 30:40.920 --> 30:49.370 So in this node that JSON defines all the key variables needed to make the call, like who does supervisor 30:49.370 --> 30:56.250 agent is calling, what the purpose of the call is, and how he should speak to the supervisor. 30:56.250 --> 31:03.290 Agent passes those variables to this node, which then triggers a separate workflow. 31:03.290 --> 31:07.330 So this one that handles the actual call on VPI. 31:07.650 --> 31:15.530 Now these fields are automatically filled by N810 at runtime based on what you ask supervisor to do. 31:15.810 --> 31:24.770 So now we are going to use those same dynamic variables in our system prompt to make our agent sound 31:24.770 --> 31:26.490 smart and personal on every call. 31:26.770 --> 31:29.090 So let's go back to our system prompt. 31:32.050 --> 31:36.090 So by using these variables in both your workflow. 31:36.570 --> 31:46.260 So in phone call agent node and your prompt you create a Are pretty dynamic and intelligent agent that 31:46.260 --> 31:49.620 responds to exactly what the user asked. 31:49.900 --> 31:56.660 So you will be speaking with and we are using a variable with a type. 31:56.660 --> 31:58.220 And the purpose of this call is. 31:59.060 --> 32:04.220 So we use the call purpose dynamic variable and so on. 32:04.660 --> 32:09.380 And as additional nodes we want to keep responses in a specific response style. 32:10.140 --> 32:13.420 We want to maintain a tone throughout the conversation. 32:13.460 --> 32:16.380 We want to keep pauses for natural flow. 32:16.980 --> 32:19.820 Start by introducing yourself and stating the purpose. 32:20.020 --> 32:21.740 Be concise to the point. 32:22.020 --> 32:27.180 If unsure about something, say and we want to use this fallback response. 32:27.500 --> 32:29.460 So when you check the workflow. 32:32.340 --> 32:38.260 You can see we want our voice agent to say something like I'm not sure, but I will check with them 32:38.260 --> 32:39.220 and get back to you. 32:39.820 --> 32:42.150 Do not mention actions outside the call. 32:42.190 --> 32:46.710 Like emails, calendar updates, etc. do not repeat yourself and be very polite and friendly. 32:47.310 --> 32:51.550 Now, just to make it clear, this is just the schema. 32:51.670 --> 32:58.590 So the template and when the workflow runs and it will automatically fill in most of these fields using 32:58.590 --> 33:04.990 data pulled from earlier steps like the contact lookup or the telegram message. 33:06.030 --> 33:14.030 But you can set defaults here, like I've done with the tone and the speaking style and the fallback 33:14.030 --> 33:14.630 response. 33:14.830 --> 33:20.390 So every time a call is made, the agent knows to speak in a friendly tone. 33:20.590 --> 33:26.910 Keep responses concise and pause briefly between sentences, while things like first name, phone number, 33:26.910 --> 33:31.510 and instructions will be dynamically filled based on the context of your request. 33:31.750 --> 33:33.070 Now we are back in Vaapi. 33:33.990 --> 33:34.870 Just briefly. 33:34.910 --> 33:38.870 Here you can change the transcriber so you can switch to a different provider. 33:39.270 --> 33:43.470 But I like Deepgram program also can change the language and model. 33:44.190 --> 33:49.790 And here invoice section you can change the provider of of the voice. 33:49.950 --> 33:55.030 So you can switch to 11 labs and others OpenAI for example. 33:55.030 --> 34:02.150 And here you can pick your favorite voice right here you can test your voice agent and talk to it to 34:02.190 --> 34:03.830 check if you like the voice, for example. 34:04.150 --> 34:10.470 Now let's actually look at how to set up a phone number in Wapi so your agent can actually make calls. 34:11.390 --> 34:13.230 So click on phone numbers. 34:16.590 --> 34:17.590 Create phone number. 34:18.870 --> 34:22.830 Now if you are based in the US you can use the free Wapi number. 34:23.430 --> 34:27.230 So just enter area code and click create. 34:27.230 --> 34:29.390 And these numbers are completely free. 34:29.390 --> 34:33.670 And you can have up to ten of them in your account. 34:33.950 --> 34:39.790 So if you are outside of the US like me, you will need to use the Import Twilio tab. 34:40.030 --> 34:42.720 So you we need to go to Twilio. 34:42.920 --> 34:47.000 Create an account, purchase a number and then import it here. 34:47.760 --> 34:54.320 So once it's connected, you need to assign it to your voice agent in the phone number section like 34:54.320 --> 34:55.360 I've done with this one. 34:55.360 --> 35:02.120 And I'm going to show you how to create an account on Twilio and purchase a number in a separate tutorial. 35:02.360 --> 35:06.280 So please go and find the tutorial in this section of the course. 35:06.320 --> 35:15.560 So once you buy the phone number on Twilio and the phone numbers in for you section, scroll down. 35:18.000 --> 35:19.360 You should see account info. 35:20.200 --> 35:22.080 Just copy this account ID. 35:26.040 --> 35:27.200 And paste it right here. 35:29.760 --> 35:33.080 The same for auth token. 35:34.400 --> 35:35.240 Copy it. 35:37.600 --> 35:38.440 Paste it here. 35:43.530 --> 35:45.170 And your telephone number. 35:49.930 --> 35:52.410 Then simply click on import from Twilio and that's it. 35:53.330 --> 35:54.170 Alright, great. 35:54.690 --> 36:02.090 Once you successfully import the number from Twilio, you have to assign your voice agent you just created 36:02.130 --> 36:02.810 right here. 36:03.050 --> 36:05.850 Alright, now let's go back to our workflow. 36:05.890 --> 36:07.050 Alright, perfect. 36:07.090 --> 36:08.850 Now we've got the full setup complete. 36:09.370 --> 36:17.570 We have built the domain and workflow, connected the three sub workflows for email, calendar and phone 36:17.570 --> 36:21.890 calls, and linked everything to our AI agent in Waapi. 36:22.050 --> 36:27.010 There is just one more important part I want to walk you through, and that's the system prompt for 36:27.010 --> 36:28.450 our supervisor agent. 36:28.690 --> 36:29.770 So let's have a look. 36:33.530 --> 36:34.570 Let's expand it. 36:36.370 --> 36:39.860 Now this prompt is what tells the supervisor agent Exactly. 36:39.900 --> 36:41.100 So the role is your. 36:41.140 --> 36:41.580 Emily. 36:42.140 --> 36:43.500 That means personal assistant. 36:43.540 --> 36:47.060 Your primary role is to delegate tasks to the appropriate tool. 36:47.380 --> 36:49.380 You don't complete tasks yourself. 36:49.380 --> 36:55.220 Instead, you ensure that each request is routed correctly with all necessary details passed to the 36:55.220 --> 36:56.740 right tool for execution. 36:57.860 --> 36:59.100 Now, we listed all the tools. 36:59.100 --> 37:05.060 So phone call agent which initiates phone calls by following provided instructions ensuring clear and 37:05.060 --> 37:08.700 effective communication contacts data. 37:08.980 --> 37:14.260 So for any task involving communication calls, email scheduling, always retrieve and verify contact 37:14.260 --> 37:18.620 details using the contact data tool before passing them to the relevant agent. 37:18.620 --> 37:25.260 So we want our supervisor agent to use contacts Data Google Sheet to retrieve all the contact details 37:25.260 --> 37:26.660 to perform all the tasks. 37:28.420 --> 37:29.260 Email agents. 37:29.260 --> 37:34.780 So use this for handling all email related actions such as sending messages, setting up replies, or 37:34.780 --> 37:35.980 for forwarding emails. 37:35.980 --> 37:40.910 Only send emails to verified email The addresses found in the contacts list. 37:40.910 --> 37:44.470 Always sign emails as Damian and never use placeholders like your name. 37:45.070 --> 37:46.590 So of course you can modify this prompt. 37:46.590 --> 37:48.630 You can change it to your name. 37:49.630 --> 37:55.150 If an email address for a recipient cannot be found in the contact list, do not attempt to send the 37:55.150 --> 38:00.990 email or use placeholder addresses like example.com calendar agent. 38:01.030 --> 38:04.270 Use this for scheduling, updating, or managing calendar events. 38:04.790 --> 38:08.430 And now very important how to use the phone call agent. 38:08.430 --> 38:12.390 If the user wants to make a phone call to get the phone call agent tool. 38:12.710 --> 38:16.430 Ensure the following details are passed to the calling agent. 38:17.070 --> 38:20.270 First name so the name of the person or business being called. 38:20.310 --> 38:29.310 Type so the contact type, instructions, phone number, call purpose, response style, tone, etc.. 38:29.310 --> 38:34.750 So these are the exact same variables we saw earlier in the phone call agent note. 38:34.990 --> 38:40.480 And that's because the supervisor agent is responsible for passing those values to the tool. 38:40.680 --> 38:42.160 When it's time to make the call. 38:42.160 --> 38:47.360 So the calling agent already understands it is a personal assistant making the call, so only provide 38:47.360 --> 38:48.120 essential details. 38:48.120 --> 38:50.360 Avoid unnecessary contacts or explanations. 38:50.680 --> 38:55.880 And here we are providing examples for all the variables. 38:59.480 --> 39:01.200 You also have rules and best practices. 39:01.440 --> 39:02.920 So always respond in English. 39:02.920 --> 39:08.000 Always delegate tasks using the correct tool and never complete them manually. 39:09.760 --> 39:10.560 And so on. 39:11.520 --> 39:18.920 And when a task requires using one or more of these tools, make sure to identify which tool is most 39:18.920 --> 39:19.640 appropriate. 39:19.840 --> 39:23.920 Pass along the relevant details and execute the actions needed to complete the task. 39:24.240 --> 39:29.160 Your goal is to be proactive, precise, and organized in managing these resources to provide a smooth 39:29.160 --> 39:30.400 experience for the user. 39:30.440 --> 39:31.920 And lastly, we also provide. 39:31.960 --> 39:33.720 Here is the current time and date. 39:34.160 --> 39:34.560 All right. 39:34.600 --> 39:35.080 Awesome. 39:35.920 --> 39:37.040 Let's close it for now.