Skip to content

Latest commit

 

History

History
102 lines (69 loc) · 8.34 KB

README.md

File metadata and controls

102 lines (69 loc) · 8.34 KB

Building a GenAI Knowledgebase from Files in SharePoint Online

Prerequisites

In order to contain all Python dependencies and avoid conflicts, it is highly recommended you use a conda environment or any other mechanism for an isolated virtual environment. You can download/install miniconda here.

Create a new environment using the configuration provided:

conda env create -f environment.yml

Configuring your App

All the configuration values that you'll use will be contained within a .env file. Copy the .env.sample file and rename it to .env. Then, open the file and paste the values for the environmental variables as you follow the rest of this doc.

Create an Azure OpenAI Resource

You'll first need to have OpenAI service made available to your organization/subscription. If that is already done, create a new Azure OpenAI resource.

Select your subscription and create a new resource group. Give the resource group a unique name. For region, select one that is closest to you, but note that some regions are more heavily utilized than others and that can hinder performance or the service can return an error that it's received too many requests at the time. Also give your Azure OpenAI resource itself a unique name. Select a pricing tier that works for you, then click Review + create.

Create Azure OpenAI resource

With the resource created, it will forward you to the overview page for your new Azure OpenAI resource. Click the Keys and Endpoint tab in the left navigation pane. Copy the value for KEY 1 and paste it into the .env file in between the quotes for the variable AZ_OAI_API_KEY. Copy and paste the Endpoint URL into the variable AZ_OAI_ENDPOINT.

Creating and Deploying OpenAI Models

With the resources created, you'll next need to deploy two different models. One to generate the vectorized embeddings that will be stored in Azure AI Search indexes and the other that will perform the human chat completions.

From the Azure OpenAI resource page, click the Model deployments tab in the left navigation (might require you to go to Azure OpenAI Studio). From there, click the Create new deployment button and select the text-embedding-ada-002 model. Give it a deployment name that will let you know this is going to be used for embeddings and then click Create. Repeat this creation process except select gpt-35-turbo for your model this time and give it a name that makes it clear it's for chat compleitions.

Deploy Azure OpenAI models

In your .env file, paste in the values for AZ_OAI_EMBEDDINGS_DEPLOYMENT_NAME and AZ_OAI_CHAT_DEPLOYMENT_NAME accordingly.

Azure AI Search Service Configuration

From your Azure Portal, go to the Azure AI Search service to start creating a new resource.

Click the Create button and fill out the form using the same resource group that you used in your Azure OpenAI service. After selecting a region, giving it a unique name, and selecting the pricing tier, click Review+create and then finish creating the AI Search resource.

Create Azure AI Search resource

Copy and paste the Url value from the resource overview page to your .env file for the variable AZ_AISEARCH_ENDPOINT value. Also fill in/paste the name you gave the search service for the variable AZ_AISEARCH_NAME. You can enter any name for the AZ_AISEARCH_INDEX_NAME variable (e.g., sharepoint-index).

Azure AI Search Service API Key

To make requests to AI Search API, you'll need API keys for this example. To get the primary admin key, click the Keys settings tab in the left menu. Copy and paste this value for the variable AZ_AISEARCH_ADMIN_API_KEY in the .env file.

Registering Your App in Microsoft Entra ID

Start by visiting App registration management in Azure Entra ID. Click the New registation button and within the form, use the display name sharepoint-doc-indexer. Click Register to register the app in your single tenent.

Register Azure Entra ID app

With the app registered, copy the Directory (tenant) ID and paste paste it into your .env file for the environmental variable AZ_TENANT_ID. Do the same for Application (client) ID for the environmental variable AZ_CLIENT_ID. You'll also need to provide this client Id to your SharePoint Administrator. They'll use this to provide access to the specified SharePoint site.

App client and tenant info

NOTE: The SharePoint administrator will need to attach and grant your app sites.selected permissions on their side first for your app/client before you can continue!

Add Permissions to Registered App

Visit the Entra ID app's API permissions to give read permissions via the Microsoft Graph API. Your SharePoint administrator will need to provide permission to the site separately. To do so,

  1. Click Add a permission.
  2. **Click Microsoft Graph. It's the very large block at the top.
  3. Select deletegated permissions.
  4. Search User.read, check the corresponding box, click the Add permissions button.
  5. Click the Grant admin consent button to enable the permissions you just added.

App permissions added

Create and Get Entra ID Credentials

In order for you app to access the Microsoft Graph API, you'll need to create and retreive credentials from Entra ID. To do so from the Certificates & secrets management page,

  1. Click on **+ New client secret".
  2. Enter a description Sharepoint content retrieval.
  3. Click Add.
  4. Copy the Value field using the copy button next to the text and paste that as the value for your AZ_CLIENT_SECRET variable value in your .env file. Ignore the Secret ID since that's not the secret itself, it's the Id of the secret to manage systematically it if you wanted to.

App secret value

SharePoint Site Configuration

Go to the SharePoint Online site that you'll be pulling the files from. The value for SP_SITE_HOSTNAME in your .env file is the domain name and the value for SP_SITE_NAME is the URL path immediately after /sites/. For example, in the URL https://asdf12345.sharepoint.com/sites/contoso/Shared%20Documents/, the hostname would be asdf12345.sharepoint.com and the site name would be contoso.

Generating and Storing Vector Embeddings in Azure AI Search

Included in this repo is a script in the indexer folder that does the following:

  1. Gets access token to SharePoint using the delegated app credentials.
  2. Queries the site you've listed in the .env file for the containing folders.
  3. Retrieves and processes the .DOCX and .PDF files in the containing folders. The files are broken up into up to 100 token chunks.
  4. The chunks are used to generate the vectorized embeddings with the OpenAI text-embedding-ada-002 model.
  5. Each text chunk, the converted embeddings, and additional metadata are stored in search indexes within Azure AI Search.

Run that script by inputting the commands:

cd app_dev
python ingest_sp_files.py

Creating and Running the RAG App w/ Azure OpenAI

The app utilizes the Prompt Flow SDK to perform a single test run of the app. The app will query your Azure OpenAI completions endpoint and provide additional context that it retrieved from Azure AI Search.

To run the app:

python app.py

Next steps

You can create variants of your system prompts and create Prompt Flow evaluations to run and test the performance of your LLM apps for different scenarios or inputs. You can also use additional tools such as an orchestrator to programatically provide additional capabilities. One such orchestrator is Semantic Kernel, which works easily with Prompt Flow.