Website AI Assistant

Library built using ML .NET, Microsoft's machine learning platform

Create and your own bespoke AI model and make predictions

AI Assistant helps visitors to your website, narrow down which of the offered products or services suits their needs.

Packages	Version	Downloads
WebsiteAIAssistant
WebsiteAIAssistant.MinimalAPI
WebsiteAIAssistant.AzureFunction
WebsiteAIAssistant.AWSLambda

Overview

Websites usually have sections that contain information about the various products or services offered by the website.

It may contain a list of the different products or services offered, along with descriptions and more information in other pages.

This AI Assistant can help visitors narrow down which of the website's products or services suits their needs,

by classifying the visitor's natural language and/or numeric based input into one of the categories of products or services offered by the website.

You can then provide more information about that category.

The API provided by the library let you create your bespoke AI model based on your training data

and then load your model and make predictions.

Example

This example is a scenario involving classifying user's natural language input to a predicted category.

Let us say a web site has Schemes, which are a types of products or services offered by the website.

The visitor's input is classified into a predicted Scheme.

Positive

Negative

Implementation

Below is the class diagram of the core component.

Helper Services

WebsiteAIAssistantCreateModelService

Below is the class diagram of the WebsiteAIAssistantCreateModelService service.

This service (& the WebsiteAIAssistantCreateModelSettings) can be wired up for dependency injection as Singleton,

and can be used to create the model and save it as a .zip file.

var createModelSettings = new WebsiteAIAssistantCreateModelSettings
{
    DataViewType = DataViewType.File,
    DataViewFilePath = Path.Combine(Environment.CurrentDirectory, "TrainingDataset.tsv"),
    AIModelFilePath = Path.Combine(Environment.CurrentDirectory, "SampleWebsite-AI-Model.zip")
};

// Register services for dependency injection
services.AddSingleton(createModelSettings);
services.AddSingleton<IWebsiteAIAssistantCreateModelService, WebsiteAIAssistantCreateModelService>();

WebsiteAIAssistantService

Below is the class diagram of the WebsiteAIAssistantService service.

The WebsiteAIAssistantService service can be used to load the model and make predictions.

Note:- If you do not explicitly load the model, the service will automatically load the model (from the specified path in the Settings) when the first prediction is made.

Note:- The WebsiteAIAssistantService service uses the PredictionEnginePool to manage the prediction engine instances,

which allows for better performance and scalability by reusing the prediction engine instances across multiple predictions.

It is also thread-safe, so it can be used in a multi-threaded environment without any issues.

The WebsiteAIAssistantService service & the WebsiteAIAssistantSettings have to be wired up for dependency injection as shown below.

services.AddWebsiteAIAssistantCore(settings =>
{
    settings.AIModelLoadFilePath = Path.Combine(Environment.CurrentDirectory, "SampleWebsite-AI-Model.zip");
    settings.NegativeConfidenceThreshold = 0.70f;
    settings.NegativeLabel = -1f;
});

Usage

If the AI model is trained using a text file, the file must be in a tab-separated format with two columns and no header.

First column contains the category Label and the other column (Feature) contains the training text data.

In the example, the training text data is all public information on the website about the various Schemes offered by the website.

Eg. ACCU Scheme has Label 0 and the training text data contains all the information about the scheme.

You can take a look at the training dataset here

Please note that the training dataset is very small and is only for demonstration purposes.

Note:- Create an entry for NegativeLabel (eg. -1 default Label, with no Feature data) in the training dataset for testing the negative case where the visitor's input does not match any of the categories of products or services offered by the website.

Using the Prediction Engine

You can set the NegativeLabel, to determine the Label for negative predictions. Default value is -1.

// Set the negative label (optional)
PredictionEngine.NegativeLabel = -10;

You can set the NegativeConfidenceThreshold to a value between 0 and 1, to determine the minimum confidence score for a negative prediction to be considered valid. Default value is 0.70.

// Set the negative confidence threshold (optional)
PredictionEngine.NegativeConfidenceThreshold = 0.50f;

Under the hood, the SdcaMaximumEntropyMulticlassTrainer is used to train the model, which is a linear classifier that optimizes the maximum entropy objective function using stochastic dual coordinate ascent (SDCA) algorithm.

You can set the options for the trainer in the SdcaMaximumEntropyOptions property of the PredictionEngine.

// Set the options for the SdcaMaximumEntropyMulticlassTrainer (optional)
PredictionEngine.SdcaMaximumEntropyOptions = new SdcaMaximumEntropyOptions
{
    BiasLearningRate = 0.1f,
    ConvergenceCheckFrequency = 10,
    MaximumNumberOfIterations = 100,
    ConvergenceTolerance = 0.01f,
    L1Regularization = 0.01f,
    L2Regularization = 0.01f,        
    Shuffle = true
};

or you can set the options for the trainer in the WebsiteAIAssistantCreateModelSettings when using the WebsiteAIAssistantCreateModelService.

Add the Nuget package or a reference to the WebsiteAIAssistant project in your ASP.NET Core application.

Step 1 :

Set the data view type and path to your training dataset or the List of ModelInput. The List can come from a database or any other source.

The training dataset is used to train the model and create the .zip file in Step 2.

PredictionEngine.DataViewType = DataViewType.File;
PredictionEngine.DataViewFilePath = "TrainingDataset.tsv";

You can also set SdcaMaximumEntropyOptions to fine tune the trainer to your needs (as shown earlier).

Step 2 :

Create the model only once and save it as a .zip file, and reuse the file for subsequent predictions.

// Path to save model
string modelPath = Path.Combine(Environment.CurrentDirectory, "SampleWebsite-AI-Model.zip");

await PredictionEngine.CreateModelAsync(modelPath);

Step 3 :

Load the .zip model file and create the prediction engine.

// Path to load model
string modelPath = Path.Combine(Environment.CurrentDirectory, "SampleWebsite-AI-Model.zip");

await PredictionEngine.LoadModelAsync(modelPath);

Step 4 :

Test the prediction engine with a sample input.

var input = new ModelInput { Feature = "What are the requisites for carbon credits?" };

var prediction = await PredictionEngine.PredictAsync(input);

Using the Helper Services

First wire up the services for dependency injection as shown in the previous section.

Step 2 :

Create the model only once and save it as a .zip file, and reuse the file for subsequent predictions.

var createModelService = _serviceProvider.GetRequiredService<IWebsiteAIAssistantCreateModelService>();

await createModelService.CreateModelAsync();

Step 3 :

Test the prediction engine with a sample input.

var aiAssistantService = _serviceProvider.GetRequiredService<IWebsiteAIAssistantService>();

var input = new ModelInput { Feature = "What are the requisites for carbon credits?" };

var prediction = await aiAssistantService.PredictAsync(input);

Logging

If you want to log the model creation, load & predictions, you can implement the IWebsiteAIAssistantLogger interface.

Then, pass an instance of your logger implementation to the PredictionEngine

PredictionEngine.Logger = new YourLoggerImplementation();

Or if you are using the helper services, you can register it in the DI container as Singleton.

services.AddSingleton<IWebsiteAIAssistantLogger, YourLoggerImplementation>();

Your logger implementation can be like shown here for ASP .NET Core.

Tests

You can browse the Tests to see how to use the library.

Sample

You can find

• the training dataset used in the samples here.

• a sample Console App using the library here.

• a sample ASP .NET Core Minimal API using the library here.

• a sample Azure Function using the library here.

• a sample AWS Lambda using the library here.

Car Category Classification Model

In this scenario, the classification into a category is done based on numeric training data too.

The Car categories

public enum CarCategory
{
    None = -1,
    TwoDoorBasic = 0,
    TwoDoorLuxury = 1,               
    FourDoorBasic = 2,
    FourDoorLuxury = 3
}

The training data

-1	
0	2 door
0	basic
0	low price $ 20,000
0	mid price $ 25,000
0	high price $ 30,000
1	2 door
1	luxury
1	low price $ 40,000
1	mid price $ 45,000
1	high price $ 50,000
2	4 door
2	basic
2	low price $ 60,000
2	mid price $ 65,000
2	high price $ 70,000
3	4 door
3	luxury
3	low price $ 80,000
3	mid price $ 85,000
3	high price $ 90,000

The Unit tests on the model

You can see the inputs to the model along with the predicted category.

[Theory]
[InlineData("price $ 42,000", CarCategory.TwoDoorLuxury)]
[InlineData("price $ 39,000", CarCategory.TwoDoorBasic)]
[InlineData("price $ 53,000", CarCategory.TwoDoorLuxury)]
[InlineData("4 door price $ 67,000", CarCategory.FourDoorBasic)]
[InlineData("luxury price $ 88,000", CarCategory.FourDoorLuxury)]
[InlineData("luxury price $ 62,000", CarCategory.TwoDoorLuxury)]
[InlineData("2 door price $ 29,000", CarCategory.TwoDoorBasic)]
[InlineData("low price $ 55,000", CarCategory.TwoDoorLuxury)]
[InlineData("high price $ 34,000", CarCategory.TwoDoorBasic)]
[InlineData("What is the colour of a rose?", CarCategory.None)]
public async Task Load_Predict_Service_CarCategory(string userInput, CarCategory expectedResult)
{
    // Arrange                      
    var aiAssistantService = _aiAssistantServiceProvider!.GetRequiredService<IWebsiteAIAssistantService>();

    var input = new ModelInput { Feature = userInput };

    // Act
    var prediction = await aiAssistantService.PredictAsync(input);

    // Assert
    Assert.NotNull(prediction);
    Assert.Equal(expectedResult, (CarCategory)prediction.PredictedLabel);
}

You can find the tests of a sample car category classification model created using the library here.

The training dataset used for this model here.

Website AI Assistant Minimal API

There is a Minimal API endpoint which you can directly use in your API project.

You create your model and save it as a .zip file as shown above, and then just provide the path to load the model in the Minimal API settings.

Read more

GET /ai/{input}

Website AI Assistant Azure Function

There is an Azure Function endpoint which you can directly use in your Functions project.

You create your model and save it as a .zip file as shown above, and then just provide the path to load the model in the settings.

Read more

GET /ai/{input}

Website AI Assistant AWS Lambda

There is an AWS Lambda endpoint which you can directly use in your Lambda project.

You create your model and save it as a .zip file as shown above, and then just provide the path to load the model in the settings.

Read more

GET /ai/{input}