Arcee Orchestra Build an Agentic Retrieval Workflow for the Energy Industry

Transcript

Hi everybody, this is Julien from Arcee. In this video, we're going to build a new workflow with Arcee Orchestra. In this example, we're going to build a workflow to answer energy-related questions. First, we're going to walk through the workflow node by node, looking at how we can retrieve energy-related information from PDF files and other documents. We're also going to run a search on Google to retrieve some articles and then use a small language model to summarize all that information into an HTML document, which we will send to the user by email. Let's get started. As usual, the workflow starts with a start node, and the input variable for this is just the user question. We're going to take this user question and use it in two ways. First, we're going to run a Google search on the question. We're going to do this through the Composio search integration, which is very simple. We just need to pass the question, and of course, we'll retrieve search results and use those articles in the report that we write. We're also going to use the question for knowledge retrieval. This is one of the important nodes available in Arcee Orchestra. As the name implies, we can use the user question to retrieve meaningful information from documents. In this case, I've attached a bunch of PDF files, and that retrieval node is going to find chunks in these documents that are related to my question. You could use this in different scenarios, such as product documentation or customer support documentation. Here, we're interested in energy, and we're just passing the question, hoping that the retrieval node will do its job. Next, we take those two outputs—the list of URLs coming from Google Search and the document chunks—and feed them to a model node. Here, I'm using a large language model. As we can see, we have three parameters: the knowledge retrieval data, the Google Search results, and the original question. My prompt is something like this: as an energy analyst, answer the question using the context and the search results, passing all the good stuff here. I'm asking for this particular structure: an analysis of the question, a detailed answer, and a list of relevant links so that I could keep diving into the problem using those articles. I want to have HTML formatting here. The last step is just this email integration step where I'm emailing myself the report. The subject will be just the question, and the body will be the output from the previous node. I'm making sure we get HTML. Of course, we can run the workflow manually just to test it. We could just enter a question here and run the test, but let's run it in the chat directly. Let's go and ask our question, and our question will be this: What is the impact of AI data centers on electricity consumption? Give me a regional breakdown. Let's see. We can see the workflow as executed, and the model is returning a summary of the answer. I should also get an email, right? So let's go to my email and see what we got there. Here's the answer I got, and we can see it is HTML. So that's a good statement. We get a first global impact answer and then answers for each part of the world. We can see that the retriever has done its job because there is information that is pretty fresh and very unlikely to be present in the model itself. That's pretty good. Then we have the collection of links retrieved by our Google search integration. Of course, we could keep reading about this. So this is a nice little example of using an Arcee Orchestra workflow to retrieve information from in-house documents, in this case, energy documents, also running Google search to expand the scope of our answer, and using a model to generate a good story and send us an email about it. That's it for this one. Hope you liked it. If you have questions, please ask your questions in the comments. And until next time, keep rocking.