Upload data and create an index

While it’s useful to see how the model handles general questions, the chatbot should ground product-specific answers in the Contoso Camping Store catalog. That means you need a retrieval-augmented generation (RAG) pattern and a searchable index.

RAG is a pattern used in AI that uses a large language model (LLM) to generate answers with your own data. When a user asks a question, the data store is searched based on user input. The user question is then combined with the matching results and sent to the LLM using a prompt (explicit instructions to an AI or machine learning model) to generate the desired answer.

For RAG to work well, we need to find a way to search and send your data to the LLM in an efficient and cost-effective manner. This process is achieved by using an index. An index is a data store that allows you to search data efficiently. An index can be optimized for LLMs by creating vectors (text data converted to number sequences using an embedding model). A good index usually has efficient search capabilities like keyword searches, semantic searches, vector searches, or a combination of these examples. This optimized RAG pattern can be illustrated as follows.

Azure AI Foundry supports project data assets and vector indexes for RAG workflows. In hub-based projects, those assets appear under Data + indexes. The exact wizard pages can vary slightly over time, but the goal is the same: upload the product files, create a vector index backed by Azure AI Search, and then attach that project index in the chat playground.

An index asset contains important information such as:

• Where your index is stored
• How to access your index
• The modes in which your index can be searched
• Whether your index has vectors
• The embedding model used for vectors

Azure AI Search is the recommended backing store for the index in this project. The classic portal documentation often shows this workflow starting from the Chat playground, but in this guided project you create the same project assets from Data + indexes so you can inspect them directly. For more information, see Hub resources overview (classic), Build and consume vector indexes in Microsoft Foundry portal (classic), and Retrieval augmented generation (RAG) and indexes.

Let’s now upload the data and then create an index.

Upload data

Data can come from an existing Azure Storage path, a URL, or an upload from your local machine. In this exercise, you upload the Products folder into the project as a data asset.

Let’s add the Contoso Camping Store product data via an upload of the products folder.

1. Within Azure AI Foundry, in the left navigation under My assets select Data + indexes.
2. Select + New data.
3. For the Data source drop-down, select Upload files/folders.
4. Select Upload files or folders > Upload folder.
5. Browse to where you saved the Products folder and select the folder for upload.
6. Select Next.
7. For the Data name field, enter: Products.
8. Select Create.

Create an index

Now that the product data is uploaded, create a project index named products-index.

1. Stay in Data + indexes and open the Indexes tab. If your portal instead opens the index wizard from the Chat playground under Add your data, use that flow - the end result is the same.
2. Select + New index.
3. When prompted for the source data, choose the Products data asset you just uploaded.
4. For index storage, choose your Azure AI Search resource. If you don't already have one connected, create or connect one now.
5. When prompted to vectorize the data, first select the Azure OpenAI connection and then choose the embedding deployment you created in the previous unit (for example, text-embedding-3-large or text-embedding-3-small). Don't select your chat deployment here - vectorization uses an embedding deployment to convert the source text into vectors. If no embedding deployment is listed, return to the previous unit and deploy one before continuing.
6. For the Index name field, enter: products-index.
7. Keep the remaining default settings unless your environment requires something different, and then select Create vector index.

Test the model with the index

We can now return to the Chat playground and validate whether the model grounds product-specific responses in the uploaded data.

1. Navigate to the Chat playground.
2. On the Chat playground page, within the Setup section, expand the Add your data section.
3. Select the option to use an existing project index, and then choose products-index.

In the chat window, test the following prompts individually to validate whether the model references the product data:

The model should now ground product-specific answers in the retrieved product data you provided instead of relying only on general model knowledge.