F
freistli
This post is going to explain how to use Advanced RAG Service easily verify proper RAG tech performance for your own data, and integrate it as a service endpoint into Copilot Studio.
This time we use CSV as a sample. CSV is text structure data, when we use basic RAG to process a multiple pages CSV file as Vector Index and perform similarity search using Nature Language on it, the grounded data is always chunked and hardly make LLM to understand the whole data picture.
For example, if we have 10,000 rows in a CSV file, when we ask "how many rows does the data contain and what's the mean value of the visits column", usually general semantic search service cannot give exact right answers if it just handles the data as unstructured. We need to use different advanced RAG method to handle the CSV data here.
Thanks to LLamaIndex Pandas Query Engine, which provides a good idea of understanding data frame data through natural language way. However to verify its performance among others and integrate to existing Enterprise environment, such as Copilot Studio or other user facing services, it definitely needs AI service developing experience and takes certain learning curve and time efforts from POC to Production.
Advanced RAG Service supports 6 latest advanced indexing techs including CSV Query Eninge, with it developers can leverage it to shorten development POC stage, and achieve Production purpose. Here is detail step to step guideline:
a. Deploy below models in [Azure OpenAI Service](Create an Azure AI services resource - Azure AI services), get its Service URL and Key:
b. Create [Azure Document Intelligent](Create a Document Intelligence (formerly Form Recognizer) resource - Azure AI services) Service, get its Service URL and Key.
c. One Docker Environment:
a. In Docker environment, run this command to clone the dockerfile and related config sample:
b. In the AdvancedRAG folder, rename .env.sample to .env
c. In the .env file, configure necessary environment variables. In this tutorial, let's configure:
NOTE:
d. Build your own docker image:
e. Run this docker:
f. Access http://localhost:8000/
a. Click the CSV Query Engine tab, upload a test CSV file, click Submit
b. Click the Chat Mode tab, now we can use Natural Language to test how good the CSV Query Engine at understanding CSV content:
The Advanced RAG Service is built with Gradio and FAST API. It opens necessary API Endpoints by default. We can turn off any of them in the .env settings.
The Chat endpoint can be used for different index types query/search. Since we are using "CSV Query Engine", now it is:
The response is:
Using this method, we can easily integrate the specific RAG capability to our own service, such as Copilot Studio. Before that, let's publish the service first.
We have different methods to release docker as an app service. Here are the generate steps when we use Azure Contain Registry and Azure Container App.
a. Create Azure Container Registry resource [ACRNAME], upload your tested docker image to it. The command is:
b. Create an Azure Container App, deploy this docker image, and deploy it. Don't forget enable Session Affinity for the Container App.
To automate the Azure Container App deployment, I provided deploy_acr_app.sh in the repo.
To use it:
Note: for more details about this sh, can refer to this guideline.
After around 7~8 minutes, the Azure Container App will be ready. You can check the output and access it directly:
To protect your container app, can follow this guide to enable authentication on it.
Enable authentication and authorization in Azure Container Apps with Microsoft Entra ID
By default, we need to upload a CSV to the AdvRAG service before analysis. The service always saves the uploaded file to its local temp folder on server side. And then we can use temp file path to start the analysis query.
To skip this step, we can save common files in subfolder rules of the AdvancedRAG folder, and then build your docker image. The files will be copy to the docker itself. As a demo, I can put a CSV file in AdvancedRAG/rules/files, and then pubish the docker to Azure.
a. Open Copilot Studio, create a new Topic, use "CSV Query" to trigger it.
b. For demo purpose, I upload a test CSV file and got its path, then put it into a variable:
c. Now let's add a Question step to ask what question the user want to ask:
d. Click "+", "Add an Action", "Create a flow". We will use this new flow to call AdvancedRAG service endpoint.
e. We need Query, File_Path, System_Message as input variables.
e. In the flow Editor, let's add an HTTP step. In the step, post the request to the AdvancedRAG endpoint as below:
Save the flow as ADVRAGSVC_CSV, and publish it.
f. Back to Copilot Studio topic, we will add the action as below, and set input variables as need:
g. Publish and open this Custom Copilot in Teams Channel based on this guide.
h. Now we can test this topic lit this, as we see, even I used gpt-4o-mini here, the response accuracy is very good:
From above, it shows how to quickly verify potential useful RAG techs (Pandas Query Engine) in the AdvancedRAG service studio, expose and publish it as REST API endpoint which can be used by other service, such as Copilot Studio.
The overall process can be applied to Knowledge Graph, GraphRAG, Tree Mode Summary and other type indexes with this AdvnacedRAG service. The main difference is for CSV, we can use its file path as the index name directly, for other indexes, we need to build index firstly, and then use index folder path as Index name. In this way developers can efficiently move from proof of concept to production, leveraging advanced RAG capabilities in their own services.
The AdvancedRAG service focuses on key logic and stability of different important index types, the efficiency to be landed into M365 AI use cases. For any feature improvement ideas, feel free to visit below repos to create issues, fork projects and create PRs.
Docker Deploy Repo: GitHub - freistli/AdvancedRAG: Efficient way to evaluate which Index type can exactly address specific LLM RAG use cases
Source Code Repo: GitHub - freistli/AdvancedRAG_SVC
Exploring the Advanced RAG (Retrieval Augmented Generation) Service
Continue reading...
This time we use CSV as a sample. CSV is text structure data, when we use basic RAG to process a multiple pages CSV file as Vector Index and perform similarity search using Nature Language on it, the grounded data is always chunked and hardly make LLM to understand the whole data picture.
For example, if we have 10,000 rows in a CSV file, when we ask "how many rows does the data contain and what's the mean value of the visits column", usually general semantic search service cannot give exact right answers if it just handles the data as unstructured. We need to use different advanced RAG method to handle the CSV data here.
Thanks to LLamaIndex Pandas Query Engine, which provides a good idea of understanding data frame data through natural language way. However to verify its performance among others and integrate to existing Enterprise environment, such as Copilot Studio or other user facing services, it definitely needs AI service developing experience and takes certain learning curve and time efforts from POC to Production.
Advanced RAG Service supports 6 latest advanced indexing techs including CSV Query Eninge, with it developers can leverage it to shorten development POC stage, and achieve Production purpose. Here is detail step to step guideline:
POC Stage
Prerequisite
a. Deploy below models in [Azure OpenAI Service](Create an Azure AI services resource - Azure AI services), get its Service URL and Key:
Code:
gpt-4o mini or gpt-4o
text-embedding-3-smallb. Create [Azure Document Intelligent](Create a Document Intelligence (formerly Form Recognizer) resource - Azure AI services) Service, get its Service URL and Key.
c. One Docker Environment:
Code:
Windows Docker, WSL2, Debian or UbuntuSetup
a. In Docker environment, run this command to clone the dockerfile and related config sample:
Code:
git clone https://github.com/freistli/AdvancedRAG.gitb. In the AdvancedRAG folder, rename .env.sample to .env
Code:
cd AdvancedRAG
mv .env.sample .envc. In the .env file, configure necessary environment variables. In this tutorial, let's configure:
Code:
#Azure OpenAI Setting
AZURE_OPENAI_API_KEY=
AZURE_OPENAI_Deployment=gpt-4o-mini
AZURE_OPENAI_EMBEDDING_Deployment=text-embedding-3-small
AZURE_OPENAI_ENDPOINT=https://[name].openai.azure.com/
# Azure Document Intellenigence
DOC_AI_BASE=https://[name].cognitiveservices.azure.com/
DOC_AI_KEY=NOTE:
Azure AI Search is not required here because the CSV Query Engine doesn't need it. If you want to try Azure AI Search Index, then configure it.d. Build your own docker image:
Code:
docker build -t dockerimage:tag .e. Run this docker:
Code:
docker run -p 8000:8000 dockerimage:tagf. Access http://localhost:8000/
Code:
Now we can see this service UIEvaluate Response Quality & Performance
a. Click the CSV Query Engine tab, upload a test CSV file, click Submit
b. Click the Chat Mode tab, now we can use Natural Language to test how good the CSV Query Engine at understanding CSV content:
Expose it as a REST API Endpoint
The Advanced RAG Service is built with Gradio and FAST API. It opens necessary API Endpoints by default. We can turn off any of them in the .env settings.
The Chat endpoint can be used for different index types query/search. Since we are using "CSV Query Engine", now it is:
Code:
POST {{LocalURL}}/advchatbot/run/chat
content-type: application/json
{
"data": [
"how many records does it have",
"",
"CSV Query Engine",
"/tmp/gradio/86262b8036b56db1a2ed40087bbc772f619d0df4/titanic_train.csv",
"You are a friendly AI Assistant" ,
false
]
}The response is:
Code:
{
"data": [
"The dataset contains a total of 891 records. If you have any more questions about the data, feel free to ask!",
null
],
"is_generating": true,
"duration": 3.148253917694092,
"average_duration": 3.148253917694092,
"render_config": null,
"changed_state_ids": []
}Using this method, we can easily integrate the specific RAG capability to our own service, such as Copilot Studio. Before that, let's publish the service first.
Publish to Azure
We have different methods to release docker as an app service. Here are the generate steps when we use Azure Contain Registry and Azure Container App.
a. Create Azure Container Registry resource [ACRNAME], upload your tested docker image to it. The command is:
Code:
az login
az account set -s [your subscription]
az acr login -n [ACRNAME]
docker push [ACRNAME].azurecr.io/dockerimage:tagb. Create an Azure Container App, deploy this docker image, and deploy it. Don't forget enable Session Affinity for the Container App.
To automate the Azure Container App deployment, I provided deploy_acr_app.sh in the repo.
Code:
#!/bin/bash
set -e
if [ $# -eq 0 ]
then
echo "No SUF_FIX supplied, it should be an integer or a short string"
docker image list
exit 1
fi
SUF_FIX=$1
RESOURCE_GROUP="rg-demo-${SUF_FIX}"
LOCATION="eastus"
ENVIRONMENT="env-demo-containerapps"
API_NAME="advrag-demo-${SUF_FIX}"
FRONTEND_NAME="advrag-ui-${SUF_FIX}"
TARGET_PORT=8000
ACR_NAME="advragdemo${SUF_FIX}"
az group create --name $RESOURCE_GROUP --location "$LOCATION"
az acr create --resource-group $RESOURCE_GROUP --name $ACR_NAME --sku Basic --admin-enabled true
az acr build --registry $ACR_NAME --image $API_NAME .
az containerapp env create --name $ENVIRONMENT --resource-group $RESOURCE_GROUP --location "$LOCATION"
az containerapp create --name $API_NAME --resource-group $RESOURCE_GROUP --environment $ENVIRONMENT --image $ACR_NAME.azurecr.io/$API_NAME --target-port $TARGET_PORT --ingress external --registry-server $ACR_NAME.azurecr.io --query properties.configuration.ingress.fqdn
az containerapp ingress sticky-sessions set -n $API_NAME -g $RESOURCE_GROUP --affinity stickyTo use it:
Code:
chmod a+x deploy_acr_azure.sh
az login
az account set -s [your subscription]
./deploy_acr_azure.sh [suffix number]Note: for more details about this sh, can refer to this guideline.
After around 7~8 minutes, the Azure Container App will be ready. You can check the output and access it directly:
To protect your container app, can follow this guide to enable authentication on it.
Enable authentication and authorization in Azure Container Apps with Microsoft Entra ID
Integrate into Copilot Studio
By default, we need to upload a CSV to the AdvRAG service before analysis. The service always saves the uploaded file to its local temp folder on server side. And then we can use temp file path to start the analysis query.
To skip this step, we can save common files in subfolder rules of the AdvancedRAG folder, and then build your docker image. The files will be copy to the docker itself. As a demo, I can put a CSV file in AdvancedRAG/rules/files, and then pubish the docker to Azure.
a. Open Copilot Studio, create a new Topic, use "CSV Query" to trigger it.
b. For demo purpose, I upload a test CSV file and got its path, then put it into a variable:
c. Now let's add a Question step to ask what question the user want to ask:
d. Click "+", "Add an Action", "Create a flow". We will use this new flow to call AdvancedRAG service endpoint.
e. We need Query, File_Path, System_Message as input variables.
e. In the flow Editor, let's add an HTTP step. In the step, post the request to the AdvancedRAG endpoint as below:
Save the flow as ADVRAGSVC_CSV, and publish it.
f. Back to Copilot Studio topic, we will add the action as below, and set input variables as need:
g. Publish and open this Custom Copilot in Teams Channel based on this guide.
h. Now we can test this topic lit this, as we see, even I used gpt-4o-mini here, the response accuracy is very good:
From above, it shows how to quickly verify potential useful RAG techs (Pandas Query Engine) in the AdvancedRAG service studio, expose and publish it as REST API endpoint which can be used by other service, such as Copilot Studio.
The overall process can be applied to Knowledge Graph, GraphRAG, Tree Mode Summary and other type indexes with this AdvnacedRAG service. The main difference is for CSV, we can use its file path as the index name directly, for other indexes, we need to build index firstly, and then use index folder path as Index name. In this way developers can efficiently move from proof of concept to production, leveraging advanced RAG capabilities in their own services.
More Information
How to improve
The AdvancedRAG service focuses on key logic and stability of different important index types, the efficiency to be landed into M365 AI use cases. For any feature improvement ideas, feel free to visit below repos to create issues, fork projects and create PRs.
Docker Deploy Repo: GitHub - freistli/AdvancedRAG: Efficient way to evaluate which Index type can exactly address specific LLM RAG use cases
Source Code Repo: GitHub - freistli/AdvancedRAG_SVC
General Introduction
Exploring the Advanced RAG (Retrieval Augmented Generation) Service
Continue reading...