An Easy Chatbot Interface
- Automatically generates the azure resources needed for the chatbot
- Uploads the pdf documents to an azure storage account
- Indexes the pdf documents with azure search (Vectorized, Hybrid and Semantic Search)
- Uses OpenAI to generate responses to user input with links to the pdf documents (citation to page)
- On click it will open the pdf document at the correct page
- Supports multiple users with different roles (roles can be used to filter the search results and generate more relevant responses)
- Supports streaming for low-latency responses
Tip
In case you use the Azure Cloud Shell, most required components are already installed.
You just need to start the Cloud Shell and do this: Install-Module -Name Az.Search
For local installation, you have to install the following components:
- PowerShell 7+ (Installation instructions)
- Azure PowerShell Az Module (More information here)
Install-Module -Name Az -AllowClobber Install-Module -Name Az.Search Connect-AzAccount # login
- Bicep CLI (Installation instructions)
- Put your pdf documents in the pdf_documents folder
- Create your users.json file (see sample-users.json) or use the example
cp sample-users.json users.json - Fine tune the system-prompt.md and system-prompt-fewshot-examples.md files
- Run the setup script ( You can check the parameters in the deployment.bicep file )
# create a resource group New-AzResourceGroup -Name "easychatbot" -Location northeurope # deploy the chatbot - you require owner role on the resource group .\iac\deployChatbot.ps1 -ResourceGroupName "easychatbot" -Location "northeurope"
- Login to the chatbot interface and chat with your pdf data
This file contains the users that can login to the chatbot interface. The file should be in the same folder as the sample-users.json file. It has the following structure:
[
{
// username should be unique
"username": "admin",
// password can be plaintext, but should be hashed.
// In this case it has to begin with 'sha256:' and then the hashed password
// (f,e, "admin" = "sha256:8c6976e5b5410415bde908bd4dee15dfb167a9c873fc4bb8a81f6f2ab448a918")
"password": "admin",
// role allowed values: see roles.json
"role": "admin"
}
]You can also hash all plaintext passwords with the following command:
.\iac\hashPasswords.ps1This file is optional and contains the roles. You can define custom filters for azure search in the filter field or use the blobPathStartsWith field to define a filter for the blob path.
It essential in case you want to filter the search results based on the user role.
[
{
"role": "admin",
"Description": "Can find all documents because there is no filter and no blobPathStartsWith defined"
},
{
"role": "limited-access-to-myfolder",
"Description": "Can access the myfolder in documents container of blob storage",
"filter": "search.ismatch('\"*.blob.core.windows.net\\/documents\\/myfolder\\/*\"', 'metadata_storage_path')"
},
{
"role": "alternative-limited-access-to-myfolder",
"Description": "Can access the myfolder in documents container of blob storage",
"blobPathStartsWith": "/myfolder/"
}
]Links to filter:
- Azure Search Filter Syntax
- search.ismatch Lucene Query Syntax
The system-prompt.md contains the dos and donts for the chatbot. What it should do, shouldn't do and how it should behave.
You are a helpful chatbot, that helps the user to find information in documents.
You refuse to talk about politics, religion or other sensitive topics. Instead, you redirect the user to your role.The system-prompt-fewshot-examples.md contains the examples for the the chatbot to better understand the user input.
- Example to Explain words:
- "blablabla" means you talk a lot of nonsense. - "wooohhhoooo" means you are very happy.
- Example to clarify expected output:
- User: Where is the best place to go skiing? System action: Search and provide answer - User: What are my latest bookings? System action: Let the user know that you can't help with that. - User: Which hotel offers suite room? System action: Search and provide answer - User: Is there a room available in Zurich on Mach 24th? System action: Let the user know that you can't help with that.
Documentation for the supported environment variables (required and optional) for the chatbot interface. You don't need to set the environment variables, since the deployment.bicep file will set everything for you.
| Variable Name | Description | Example |
|---|---|---|
| CHATBOT_SECRET_KEY | Required, Secret Key for the chatbot interface (used for user login cookie) | keepItSecretAndDoNotTellAnyone |
| CHATBOT_STREAMING | Optional, Enable or disable streaming (default: true) | false |
| AZURE_STORAGEBLOB_CONNECTIONSTRING | Required for storage account key auth (if not using AZURE_STORAGEBLOB_RESOURCEENDPOINT) | DefaultEndpointsProtocol=https;AccountName=your_account_name;AccountKey=your_account_key;EndpointSuffix=core.windows.net |
| AZURE_STORAGEBLOB_RESOURCEENDPOINT | Required for default credential Entra ID auth (if not using AZURE_STORAGEBLOB_CONNECTIONSTRING) | https://your_account_name.blob.core.windows.net |
| AZURE_STORAGEBLOB_CONTAINER | Optional Azure Storage Blob Container Name (Default: documents) | documents |
| OPENAI_API_BASE | Required, OpenAI API Base URL | https://myazureopenainame.openai.com |
| OPENAI_API_KEY | Optional, if not set will use default credential Entra ID auth | your_openai_api_key |
| OPENAI_DEPLOYMENT_NAME | Optional, default is 'gpt-4o' | gpt-4o |
| OPENAI_EMBEDDING_DEPLOYMENT_NAME | Optional, default is 'text-embedding-ada-002' | text-embedding-ada-002 |
| AZURESEARCH_API_BASE | Required, Azure Search API Base URL | https://myazuresearchname.search.windows.net |
| AZURESEARCH_API_KEY | Optional, if not set will use managed identity of open ai service | your_azuresearch_api_key |
| AZURESEARCH_INDEX_NAME | Optional, default is 'documents' | documents |
Summarized:
- exactly one of the following variables is required:
- AZURE_STORAGEBLOB_CONNECTIONSTRING
- AZURE_STORAGEBLOB_RESOURCEENDPOINT
- and all of the following variables are required:
- CHATBOT_SECRET_KEY
- OPENAI_API_BASE
- AZURESEARCH_API_BASE
Useful links: