Aisle Platform Demo

a quick video of what this is, how it works. I'm going to make it uh pretty fast. You've seen this as I said. Uh so coming on here as a knowledge worker, I have a chat client. It is multimodel.

I can pick claude. I can pick GPT, uh Llama, DeepSeek, uh Gemini. Basically, anything that's anthropic or GPT is supported as well as anything that's on open router. In this case, I'm going to ask for claude 35 sonnet. I'm just going to say, you know, tell me 10 words about tennis, for example, so we can see it's real.

this chat client largely fully featured. You can upload files uh anything like that. Uh we can also come in and share the chat uh either with a public link or with specific people on our team if we wanted to. Uh and obviously all of this invite flow works as we would expect. Uh the other neat feature here is we have this uh fork button here.

So I'm often very deep into a research thread. I want to explore two or three options without funking up my context. uh any single system message we can come back and or machine generated message we can come back and fork at any moment which is great and this is already pretty cool we have a multi-feature chat client or multimodel chat client we can distribute to the team and not be paying for separate claw GBT Gemini licenses when we need them now if you do this type of work all the time one of the things that you run into is that you end up needing to build these prompts that you're running all the time they're just you composable prompts and we're replacing different pieces inside of them. Um, you know, it might be something that, you know, goes through and summarizes news, whatever. So, we compress this dashboard section here.

Uh, and we're in prompt management already. So, what we're going to do is build a very quick prompt together. Uh, we're going to call this really, really bad business advice. And we might just say, you know, give me some horrible business advice as a Wall Street Bets bro, for example. And you can see how we're basically building a prompt that we're going to use over and over.

Uh we can pick different models, different temperature settings, everything you would expect here. Um we can attach files within the prompt that are static. I use this all the time when I'm building say a prompt that generates me components and I'm you know attaching screenshots of you know things that we like or we can allow a file to be added at runtime. Uh before getting into that what we should do is create a parameter. So what we can say here is the business ID is business ID and if you notice we've now created a parameter in here which is something we're going to continue replacing at runtime every time we execute this prompt.

Uh we could test it out with this run button. I'm not going to wait for that to finish. That's also where if we click allow additional files, we're allowing the user to add additional files at runtime, which is a common thing we'll do for, you know, go through and analyze this financial report and pull out these certain details. A lot of common reuse for stuff like that. Now, we're going to save this on 37 sonnet.

Every prompt we create has a set of logs. Uh you can see the old versions. You can actually publish it to an endpoint uh which will require an access key you've generated. Uh which is great because now you can call these prompts uh from application code you're writing but have a nice visual place to build the prompts, manage the prompts. If you need to change an underlying model or setting, you can visually see what's going on.

More importantly, we have this ability now to come in and allow other people in the organization to view or share that prompt. We can do that only with certain people if we wanted to and that would work as well. Or what we can do is enable this prompt for usage in everybody's chat or in this case only my chat. And I'm going to come back and reload my chat now. And we'll see in this template section really bad business advice is available.

So I might give the business idea of train headlights for example. And we'll see this is immediately streaming through. So now we have this recomposable prompt that we can distribute down to chat that frankly less technical people on a team can come in and build these reusable components. they need all the time. Um, that's already really cool.

To keep this video short, I'm not going to go into absolutely everything uh in detail. There's a playground here. We can go and cross compare prompts including new ones. Uh there is this scripting which we'll go and touch in a second. Uh these are olders Ruby or Python functions that run on Lambda.

You can access the Slack SDK, uh AWS SDK. A couple of default ones are installed. There's no restrictions on external HTTP calls. This is your way to build integrations which we'll see in a second because that brings us to this workflows section. And if you build any of these agentic things uh in production, what you end up or the insight everybody comes to is you end up needing to stack many prompts together that do a single thing and chain context in between them.

So that's how we build workflows here. So we might build a basic business workflow for example. Uh, and just to show it's real, what we might do is go in and we'll get our prompt here, which was the uh, terrible business advice. And we want to run that prompt. And as we're running that prompt, we can see we had an input parameter here called business topic.

And is that a static or reference? Static just means, you know, the value is going to be the same. Reference means it's coming from another node. This is how we're managing context. In this case, it's the first node.

What we need to do is allow the user to type that in as a start parameter. And what we can do is set that to a reference. Uh, and as we're playing around, we can actually run this here. And I might call this uh train horns, for example, just to do something different. And we can use this run mode to debug workflows we're building and see, you know, what's going on with them in real time.

Uh, because I'm trying to do this fast, I'm not going to wait for the entire prompt to execute. It does. You see it's come through here. In case you thought I was trying to hide from a bug there, I'm not. Um, this is all really cool.

Workflows are usually a little bit more complex. So, if we go into something that's a little bit more real, you know what we might do is take that script we built that returns in our case an array of stock tickers. Uh, this is where scripts your opportunity for ingress egress, right? If you need to go through and pull a rag database to build a rag agent, pull internal inventory numbers, APIs, anything like that, put data. There's no restriction on what you can do in these scripts.

For now, what I'm going to do is go through and we're going to do a little bit of a variable scoping hack just so you can see that you can create variables here, pluck values, convert things to arrays, anything like that. For now, I'm just going to create a horrible array called um let's call this uh all stock advice, for example, and we're going to call it that. And we're not going to put anything inside of it right now. But then what I'm going to do is use this while loop uh map and reduce coming soon. For now, what I'm going to do is iterate the collection of stock tickers that we had received.

We know that's an array. I'm going to call each of them stock just for cleanliness. And within the loop body, what I want to do is go through and run a prompt that I had built earlier, which gives advice on a stock as a Wall Street Betspir. And what I want to pass in as the stock here is a reference value to the item inside of the while loop, which is great. So that's all good.

Uh, and what I want to do here, just to show off the variable scoping, is then what I want to do is this is going to magically convert it to an array because we expect less technical people inside of this flow. So what I'm going to do is for all of the stock advice, I want to push into that an array the output of this step here, which would be Y A if stock advice, for example. Um, and that's all, you know, pretty good. Uh, what we can do here is run this again. I'm not going to let it all finish.

We see we still need that business idea, which is fine. And, you know, you see the run mood is going to go through. It's going to iterate through each of these items one at a time. Again, in the sake of speed, I'm not going to go through that. Um, okay.

Well, one just came up, so we can see it's already gone through Nvidia. It's going to push those to the array, for example. Um, just to kind of show off how these workflows get extended, uh, as the last interesting part here, uh, what we might do is after that loop's completed, we might go back to what we had before, which is to go through and get some terrible business advice, uh, on the business topic that the user has passed in from the start parameter. And that's the node we're going to run next. This is obviously a very artificial flow.

Uh, and actually before that, what we might want to do is add this node called a chat action, which lets us add a link modal or accordion. In this case, I'm going to be fast and just add a modal called stock advice to the chat that we're going to generate. And that is going to be a reference to all the stock advice that we've received. And then after that, what I'm going to do, this never works when I'm on video at the moment because I have my zooms all messed up. There we go.

Is I'm going to go through and connect this to terrible business advice. And right now, workflows, we're building a specific return node. Right now, they return the very last value that they receive. So, this workflow is actually going to return terrible business advice here. You may have noticed some other node types here.

There are conditions. uh there is the ability to fetch content from a URL and you'll get a string version of whatever is in that URL. Uh what we can do here now is with this chat we might decide to enable this workflow inside of chat. Uh and this is what happened right we built these agents for customers and you know like these are great how do we deploy them so our less technical team can use them and oh no we're screwed. So what we can do is come back here.

I'm going to reload this and go back to my chat. And we see our business workflow is now available to run. Uh, as it's running, you actually see a status update as a user down here. And this is really cool. Right now, we have a way to distribute all of these to the team.

While this is running, just because, you know, it's obviously a few steps and it takes a while. Um, one of the things we're thinking about with these workflows specifically is it's, you know, as much as we're enabling less technical people to come in and build the prompts they need to get their work done, there's two cases of workflows. There's the ones we're deploying in organizations that are very complex agent reasoning, there's another view where, you know, we're building some basic building blocks like fetch data from rag, do x, do y, and some basic prompts that a lot of analysts are snapping together in the order they need with a couple of custom prompts before or after, which works out really well. Uh what I'm going to do very quickly so we don't go into too much detail is you may have noticed this folders section here uh as an example I have like a really good uh product summary of what is and what does it do and these are all markdown documents with metadata. Uh that's what this database node here does.

We can go through and update any of these documents which is an easy way to manage context in between calls. And if you're curious uh that's what this save button down here does as well. If we have a good chat and we want to put that in one of the databases to feed to an LLM, append it to a document, make a new document, whatever, we can do that here and just accumulate the demo. We come in, hooray, our workflow has responded. Here we see the business advice on train headlights because that is what the last node was when we decided to run that.

Even though we went through this array structure here, but we have a nice modal that's popped in. Uh I didn't do the accordion unfortunately. just wanted to go fast. Uh, which shows everything that was in that stock advice. And you'll see Nvidia and Microsoft.

It is sitting here unforatted right now. Uh, very fast demo. No doubt you're going to run into some bugs. No doubt you're going to have some questions. Uh, this is largely what it does and how you can play around.

Enjoy.