Watch PromptQL in Action

Hey folks, here are a few quick demos of PromptQL in action. To recap, PromptQL is an approach for building highly reliable AI applications on enterprise data. So the first example is on a telecom company's data. I want to know which customer is impacting the capacity of my network the most and when did they have the highest usage. Who are the top 5 customers that are impacting the capacity of my network the most? For the top one, tell me which state did they have the most usage? So when I ask this question, PromptQL first understands the question and looks at the underlying data landscape and comes up with a query plan of how to approach this problem. It tells me what the query plan is and implements a task-specific agent under the hood to do exactly that. So it goes through these multiple steps, implementing it a new agent on the fly and understands, okay, these are the top customers who are impacting the network traffic. Then it analyzes the network performance metrics of that top customer to understand the impact on the network. And then finally gives me a report that these are top 5 customers by network data usage. And specifically, Daniel Lane is the top customer that is impacting network traffic and then gives me insights on their usage. The second example is an anti-money laundering use case for a financial firm. I have received intelligence about potential trade-based money laundering, and we need to analyze the customer's transaction patterns, including suspicious currency conversions and amounts. So we have received intelligence about potential trade-based money laundering. Can you analyze the transaction patterns where there are frequent currency conversions between the sender and the receiver accounts, especially where the payment currency differs from the received currency and the transaction amounts don't align with the customer's expected behaviour? So again, PromptQL looks at the underlying data landscape and then comes up with a plan of how to approach this problem. It says that I need to find these transactions where payment currency is different from the received currency, then it joins this data with the customer information, and then it analyzes transaction patterns to identify frequent currency conversions, transaction amounts being unusual, or patterns that might indicate trade-based money laundering. and then gives me a report that says that okay there are 39 suspicious currency conversion transactions. One account specifically by Thomas Williams stands out with 4 currency conversion transactions totalling over ten thousand dollars. Several high-risk customers are engaging in cross border currency conversions and multiple transactions involve politically exposed persons or blacklisted customers. So it goes through our entire data and comes up with these insights. This final example is for a typical enterprise software company who wants to see if their highest value customer is at a risk of churn. It is a little tricky because the data is messed up and is in different systems. So I ask, find the highest bill organization that we serve. The organization ID data is messed up, so use the user's email domains to find the unique organizations. Then for this organization, fetch all the support tickets across all of their users from Zendesk. Then for each ticket thread, which means ticket details plus comments on those tickets, summarize the ticket thread. Then use these summaries to extract this organization's sentiment towards our product, what is going well, what is not going well. And if the sentiment is positive, issue a thousand dollars to their highest individually built users, most used project. If the sentiment is neutral then issue two thousand five hundred dollars and if the sentiment is negative then issue five thousand dollars. So prompt drill takes your query, looks at the super graph under the hood, comes up with a query plan, breaks it down into steps. Say first it needs to find this highest billed organization and organizations are found based on email domains of the users. And it looks at the invoice items to calculate the total billing for each organization. Then it realizes Williams.com is our highest billed organization. So now it's fetching all the support tickets and comments. by all the users of Williams.com and it's fetching this from Zendesk. And now it's going to analyze these support threads to understand the sentiments. So it's going to call another LLM under the hood for each ticket and ask it to summarize the ticket thread and then combine all of these summaries and then call another LLM to extract the sentiment and what's going well, what's not going well for this organization. And now it's asking me to confirm whether or not to issue this refund. The data delivery network under the hood says that your AI is about to call this command, and these are the parameters it's passing. Are you OK with that? So if I click on Approve, the DDN layer will allow this to happen, and the refund will get issued. So that was PromptQL in action. You can read more about PromptQL by visiting promql.hasura.io.