Hello and welcome. In this demo, we're going to build a lightweight RAG and agentic workflow assistant for a study abroad office. Here's what we'll cover.

First, we'll load and chunk the documents. Next we'll embed an index using the open-source Facebook AI Similarity Search. Then, we'll build a simple agentic graph. And finally, we'll ask a student-style question and inspect the grounded results.

By the end of this, you'll recognize the features of a minimal RAG pipeline, including citations, chunking, and retrieval parameters, and using agentic workflows. So, let's get started.

First, we'll set up our environment. Here we install the required packages. Most of these are related to LangChain. The pypdf module handles document loading. This module here is the Facebook AI Similarity Search, and this will serve as the vector database.

And finally, we have tiktoken, which is a module-handling tokenization. This cell here suppresses warnings and defines helper functions for text normalization and for previewing. For this cell here, I've configured the API keys off camera.

We're going to import documents related to studying abroad into vector storage. We'll use three PDF files to serve as our source documents. Running this cell lists the PDF files found in the current directory.

We have Handbook 1 covering visa and compliance. We have Handbook 2 covering housing and health, and Handbook 3 including travel and insurance. Now that we have the file paths, we can load these as document objects.

Once loaded, we break the documents into chunks. Here we define the size of the chunk and the number of overlap tokens. This overlap helps preserve context across boundaries.

Notice that three PDF files produced 25 document chunks after being split. Metadata is attached to each, including the file name and page number. This metadata will later be used for citations. Looking carefully at the first five document chunks, you'll notice the text overlap that we specified.

The next cell creates objects for interacting with the gpt-4o-mini model. This text-embedding-3-small model is used for embeddings.

Our vector store here is initialized with those 25 document chunks using the embedder to produce the vectors. And finally for this cell, the document retriever is configured to return the top four most relevant chunks per query.

The next cell creates the RAG state class which will be used to store model input, output, and retrieved documents. This retrieve function fetches the most relevant documents from the vector storage.

These functions here with the preceding underscores are used for formatting prompts to include the documents as context, including citations. The generate function is responsible for sending the prompts, system, and user to the model.

So far, we have the building blocks with two key steps: retrieval and generation. So, next we need to wire these steps into a state graph. Retrieval is the entry point followed by generation. The generated image demonstrates our final workflow.

With all this in place, we can test the system with realistic student queries. This ask model function runs through this workflow, returning the results as a RAG state object. This list here has several questions.

Notice the results of these questions provide answers and citations. Because these responses are grounded in actual documents, it greatly increases the odds of accurate responses. Let's check one of these citations for accuracy.

The first question is, "How long can I stay on a short stay visa and can I work during term time?" Here we see that this is sourced from the Handbook 1, page 1. Based on these two sections here, we can see that the answer is correct. We have a maximum of 90 days and up to 20 hours during an active teaching period.

This demo shows how to combine retrieval and agentic reasoning to build more accurate assistance. In this scenario, the agent doesn't have to guess. It retrieves, evaluates, and sites.

When designing RAG and agentic systems, keep chunks small and well indexed, use clear prompts with citation requirements, add guardrails for fallback behaviors, and finally, expose that retrieved context for transparency.

All right, that's going to wrap up this demo. Thanks for watching.

Lab Demo

About