2. Airbnb SMS Agent Updates 🏡 - SCAI017
Sloup
1 follower
8 Video Views·Sep 15, 2025
Nate Herk - AI Automation Society Plus\10. Client Builds (Real Projects)\2. AirBnB Booking Agent\2. Airbnb SMS Agent Updates 🏡 - SCAI017
Hey guys, so back with some more updates on the Airbnb agent.
This is today I'm just going to be looking at sort of the SMS agent part of things.
So this is kind of what we what we have right now.
So basically like it gets the it gets the text and then it puts it into the Pinecone Vector Store.
Of course, we have to just map the input.
So basically what we're looking at here is let me just go to
the example execution that I want to look at in this case.
We'll go to this one right here.
So, you know, the input comes in.
It's grabbing, you know, sort of like a query that I asked.
So I texted it.
What is January availability like?
It grabs my name, the phone number, and then we're just basically mapping stuff so that
it correctly goes into the Pinecone Vector Store as well as feeds into the agent correctly
so they can process what it's looking for.
In this case, it was just looking for availability.
So it didn't have to hit the quoter.
Normally we need information about like what dates exactly and how many people are going
to be staying.
And that is usually, you know, first of all, it's going to hit the availability check to
make sure that the dates are available and then it will go hit the quoter if, you know,
people are specified, that sort of thing.
So then obviously it's going to, you know, it will check some past information.
It has a knowledge base if there's any questions about the company or the different properties.
And then it's going to hit the SMS specialist.
So in the SMS specialist, we'll go to the exact same execution, which would have been
this one, I believe.
So all this is going to do is it's also going to check past conversations received and
conversations sent.
That way it just has context of, you know, we're dealing with this specific client.
So we want to make sure the information makes sense.
And as you can see, sort of like this is what's coming in.
It's the query, which is the, at this point, it has already done the availability check.
So as you can see, thanks for reaching out about our availability in January.
We're happy to inform you that January is fully available, blah, blah, blah.
And then what we have it doing now is it's scraping, feel free to check out the calendar
here.
So it's scraping a screenshot of the actual calendar from the booking website so that
they can really see like what dates are fully available.
So this is getting the screenshot link.
And so then here, what we wanted to do was parse out the actual screenshot.
So it's all coming in one block.
And basically I just prompted it to, you know, take the URLs that follow the screenshot one
in the case where there's two screenshots.
If it was like, hey, can we stay from January, you know, 31st until February 5th, then it's
going to want to pull January and February screenshots.
So right here, you can see screenshot two is null, but here's screenshot one.
So yeah, we're basically giving it this information to parse out, to give it to the specialist.
That way this guy knows, you know, what's coming in.
And what we wanted to do was make sure that we're mapping the screenshot to the actual
send SMS tool.
So I don't know.
I talked about this in a previous, you know, update video about this agent where I think
what I want to do is after this agent grabs like the conversation sent and received, I'd
rather have the craft SMS sort of go here after the agent and then the send SMS go also
here because there's no point for the agent to call these two tools when it's going to
happen every single time.
So yeah, working on implementing those changes, but I'm also, I'm not the only one working
on this project.
I was kind of thrown into this one halfway through.
So that's why when it comes to stuff like over here where we're looking at the quota
and the availability check, a lot of this stuff actually is taking place in a different
scenario.
We're calling, we're calling something in make and make is going through and grabbing
the quota and grabbing the scraping to get the availability.
And I don't know, it's sort of, it's interesting because like I said, I was thrown into this
project and a lot of this stuff is taking place through go high level.
So like the send SMS is actually taking place within go high level as well.
So I'll show you that in a sec.
But anyways in the send SMS, we are mapping the screenshot value.
So you can see here we have screenshot one and screenshot two that will be coming through.
And in this case, there's only one screenshot.
So this is here and this is no, but that is something that we're making sure that always
goes into the send SMS tool.
It's just easier to map it out here as an extra workflow input rather than start trying
to prompt this agent to, okay, when you're going to send SMS, always make sure the screenshot
link is included because it doesn't need that screenshot link to craft the SMS.
It just needs it in the actual parameters of sending SMS.
So let me get into the send SMS tool now.
It will make more sense.
So I know this is kind of messy.
But when I was testing out this kind of stuff, we have these three different HTTP requests
that go to send SMS through go high level.
And the thing with these is that we put in these parameters here of like screenshot one,
screenshot two.
And in the case where screenshot two was no, this, this whole HTTP request just didn't
work because it had a no field.
And sometimes it didn't even come back as an actual no, it was just a string that said
no.
So to, you know, sort of work around this, we just had these different switch or sorry,
ifs are basically here, you know, if they're both if one or the other one is no, and it's
going to go up here and it's going to go here.
So as you can see, this parameter is just going to be the actual message, the text message
and the phone number to send it to.
This one is going to be the text message, the phone number and the screenshot.
So as you can see, in this case, we got the text back, which was, you know, hi, Nate,
thanks for confirming.
Would you like to proceed blah, blah, blah.
And then we get the screenshot.
So actually, let me pull up this exact.
I have this one pulled up on my phone.
So I don't know how well you can see this.
But this is me asking, like, you know, what is what is January availability like?
And then this is sorry, this is the screenshot and then the text.
So that's that's kind of an awful way to see it.
But that's kind of what it looks like of the screenshot of the calendar coming back.
Yeah, so that is what happened in this exact execution.
And then if we look at like this one, I think I tested out doing two screenshots.
So in this case, it goes up here.
We asked for I think in this case, we asked for January and December or something about
that or December and January.
So we got two screenshots here and it gave us, you know, the two different calendar links.
But as you can see, that's how the SMS is being sent is through these screenshots.
And then the same logic takes place after all of these, which is just updating the notes
within GoHighLevel for the client to be able to see the interactions that are taking place.
And just like in case it's it doesn't send, it's going to send out an error email.
But yeah, let's see what else I think would be important to talk about here.
This is still a work in progress.
So everything's optimized as far as the best way to do it.
Like I said, this kind of stuff.
And also, we have been experimenting with, you know, the conversations here,
conversation data is here, but it's also here.
And the reason we did this is because there are certain cases where they want to look
at conversation history before it goes to send the message.
So it's kind of just like a double check.
For example, we get a query in here where, let's say I was talking to this agent and I said,
what is your availability like for, you know, January 15th through 16th or something.
And then it's going to grab that availability.
It's going to, you know, it obviously updates the database every time after we sent the message.
But it'll grab the availability and then it will respond to us with what that is.
Then I was like, okay, if I said, okay, great.
What about the 25th through the 30th?
And then that way it can check here and see that we just talked about January.
So then it knows that now we're talking about 25th through 30th also of January.
So it will check availability.
So that's why we have it here as well.
I'm still trying to figure out maybe a better way to do it as far as having everything be
more streamlined as if like, you know, it's kind of a process.
Typically the first check is to go grab availability.
And then it's to usually go grab a quote.
And sometimes what happens though is, you know, there's not enough information for
a quote to be generated.
Like we need the, obviously the data has to be available, but we also need to see
the people or how many people are coming.
So let me just grab real quick the quoter tool.
And I can show you like an example of a check that we have to make sure that we're getting
all the information we have.
So like what's going to come in is like, here's an example.
Can we check in October 10th through the 25th?
It would be me and two kids.
This sort of just, you know, parses the information.
We can actually look at a quick execution of this.
Let's go to one that succeeded.
Yeah, sure.
This one sort of said, generated a quote for James Day from January 17th to the 20th
for two guests.
They're both adults.
It's going to parse that information into check-in date, checkout date, number of guests,
number of adults.
And this is, we're formatting it like this.
Obviously we prompted it to put it like this in the JSON because we're parsing it through,
or sorry, we're passing it into an HTTP request that goes and scrapes the quote and make.
But here we basically just have like, if we need all of these three to not be empty.
So check-in date, check-out date, number of guests.
And we also put into the open AI prompts before this, if adults isn't specified, then it's
going to be the same as the number of guests.
So we just need these to all be there.
Otherwise the quoter tool would run and it would fail.
And so if those aren't filled out, then it's going to go back to the agent and say, you
know, we can't pull the quote.
We need check-in date, check-out dates.
Please try again.
Or please, please text back the client asking them for this information.
So that's sort of what goes on here.
And then it would loop back.
It would send another text.
And then once it gets the information, then it could try to run this quote.
And then here it's just obviously running the quote.
It grabs the actual price.
And then over here, it's going to respond.
So like, here's the response.
$28.59 before taxes.
So it takes that response.
It grabs it back from here.
So this is an example where, let's see, maybe this one where they would have pulled the
quote.
So yeah, in this case, they already knew it was available.
They got the quote there.
Let's see this one.
Okay.
So anyways, it grabs the quote.
It gives that to the agent, of course.
And then the agent has all the information it needs to hit the SMS specialist to go craft
that SMS and ultimately send it off.
So yeah, the email one is very similar.
It's structured a little differently because we're not doing it based on a webhook trigger
where we upload it into the conversations received.
This one is a different workflow that's just an email trigger that filters through everything
that we get.
And then again, we map it.
I'm still learning the best way to put different types of information into vector stores.
Actually, should be coming out with a little loom video for you guys here that I found
this.
I've been exploring with Quadrant, and they've got some cool ways that you can actually visualize
vector information and really cool ways to filter through stuff.
So I'll just show you guys what I've been playing around with from there, but I'm still
trying to learn the best way to take different types of data, put it in there so that there's
still a contextual context, I guess.
I don't know.
But yeah, so the email one's a little different.
I'll definitely make some more videos about that one as soon as we get that stuff more
streamlined.
But yeah, the knowledge base, obviously, we have these different tools.
Yeah, it's a straightforward idea, but then when you get into actually building it, it
gets pretty complex.
So this is probably, I don't know, maybe like the fifth iteration of this as far as
refining.
If you guys really quick want to look at the system prompt that we've got in here, it's
basically, we define the role.
You're a professional assistant for helping deal with SMS inquiries.
We gave it the context.
Your objective is to take incoming SMS messages from clients, call the correct tools.
Your job here is not to communicate with clients.
We just want you to pass information around as far as you get the availability.
You're going to pass that to the quote tool to grab the quote.
You're going to take that quote and the availability, and then you're going to pass it to
SMS specialists.
So your job is only to just call tools.
So I'd also played around with maybe doing a plan and execute agent.
But it's not exactly the same steps every time.
So it wasn't working exactly as I wanted it to.
So the tools agent was working better here.
But anyways, you know, the instructions were to, you know, review guest information.
Oh, sorry.
Using the conversation history, look at availability, look at quote.
Then obviously update the database every time a message is sent.
If the quota tool fails, it's because we need more information.
So then you would use the SMS specialist to reach out and request more information.
We gave it the tools, just outline when to use each one pretty much.
And then we just gave it some example scenarios of, you know, how it would act.
And then additional notes.
Something down here is originally it was cutting off the screenshot links.
So it would end it after .png.
But there were some more parameters after the screenshot links, and we needed that.
So down here at the notes, you know, we just made things bold.
When sending these links, do not shorten them.
Send the line, the links exactly as you receive them.
If you change the link, the automation will break.
We could probably get rid of this, actually,
ensure all messages are personalized and professional,
because we have this prompted into the crafting SMS tool.
But the goal really is just to provide human-like service
that assists the clients in planning their stay.
Yeah, so that's kind of how the prompting works in here.
Yeah, so this was just a little look at sort of, you know,
how I was working out this build and some of the issues that we had when,
you know, passing data between each of these tools.
If you guys got any questions about what I did here,
if you have any things that stick out, you're like,
oh, Nate, I feel like you could definitely make this tool a lot better,
or you could maybe combine this kind of stuff.
Please let me know down below.
Well, I'd love to talk about, you know, any ideas about this build.
But please, yeah, reach out with any other questions.
And thanks for watching this video, guys.
