Add evaluation client

[TechPreacher]

This PR adds an evaluation client that can evaluate answer correctness vs. a ground truth and document retrieval for RAG based Copilot Studio agents.

Sample Evaluation

Evaluation Dataset.csv

Name: Asus Zenbook Duo 2024
Test Type: Response Match
Test Utterance: How can I charge the removable keyboard of the Asus Zenbook Duo 2024?
Expected Response: To charge the removable keyboard of the Asus Zenbook Duo 2024 insert the bundled power adapter into the USB-C port on the keyboard or dock it with the laptop to charge via the built-in connector.
Sources: https://tenant.sharepoint.com/sample/Document%20Library%201/2.pdf;https://tenant.sharepoint.com/sample/Document%20Library%201/3.pdf

Evaluation Dataset Results.csv

Name: Asus Zenbook Duo 2024
Test Type: Response Match
est Utterance: How can I charge the removable keyboard of the Asus Zenbook Duo 2024?
Expected Response: To charge the removable keyboard of the Asus Zenbook Duo 2024 insert the bundled power adapter into the USB-C port on the keyboard or dock it with the laptop to charge via the built-in connector.
Sources: https://tenant.sharepoint.com/sample/Document%20Library%201/2.pdf;https://tenant.sharepoint.com/sample/Document%20Library%201/3.pdf
Agent Response: 
To charge the removable keyboard of the Asus Zenbook Duo 2024 you can follow these steps:
1. Connect the AC power cord to the AC/DC adapter.
2. Connect the DC power connector into your keyboard’s power (DC) input port.
3. Plug the AC power adapter into a 100V~240V power source [1].
[1]: https://tenant.sharepoint.com/sample/Document%20Library%201/2.pdf
Answer Score: 65
Sources Score: 1/2

Explanation

This RAG agent was created by adding a SharePoint knowledge source to the agent. The evaluator client asks the agent to answer the question provided in the "Test Utterance" field and compares the answer with the "Expected Response" answer.

Based on how well the agent's answer matches the expected response semantically, a score is given between 0 and 100, 10 being the worst and 100 the perfect answer. The value is stored in the "Answer Score" field. The "Sources" field provided 2 URLs the agent should have used to answer the question.

The evaluator client checks if the agent used the provided URLs to answer the question by returning them as reference links and returns it as x/y where x are the number of links provided by the agent
over y which represents the number of links expected.

[svandenhoven]

The EvalClient app runs when isolated from the other samples. When the .sln of the whole solution is opened it gives error

Projects that use central package version management should not define the version on the PackageReference items but on the PackageVersion items: Azure.Identity;Microsoft.Extensions.Configuration.UserSecrets;Microsoft.Extensions.AI.OpenAI;Microsoft.Agents.Client;Microsoft.Extensions.Hosting;Microsoft.Identity.Client;CsvHelper;Microsoft.Agents.CopilotStudio.Client;Microsoft.Identity.Client.Extensions.Msal;Azure.AI.OpenAI;Microsoft.Agents.Authentication.

[tracyboehrer]

@TechPreacher This is great! We will take some action on this after the holidays. Long term, before the GA, we will handle samples differently. Namely using released package versions instead of being part of the SDK solution and using project references. We aren't quite ready to make that change, so what you are doing is the correct thing for now.

[TechPreacher]

The EvalClient app runs when isolated from the other samples. When the .sln of the whole solution is opened it gives error

Projects that use central package version management should not define the version on the PackageReference items but on the PackageVersion items: Azure.Identity;Microsoft.Extensions.Configuration.UserSecrets;Microsoft.Extensions.AI.OpenAI;Microsoft.Agents.Client;Microsoft.Extensions.Hosting;Microsoft.Identity.Client;CsvHelper;Microsoft.Agents.CopilotStudio.Client;Microsoft.Identity.Client.Extensions.Msal;Azure.AI.OpenAI;Microsoft.Agents.Authentication.

This is fixed now. Thanks for noticing!

[TechPreacher]

@TechPreacher This is great! We will take some action on this after the holidays. Long term, before the GA, we will handle samples differently. Namely using released package versions instead of being part of the SDK solution and using project references. We aren't quite ready to make that change, so what you are doing is the correct thing for now.

Thanks @tracyboehrer ! You can always reach me internally at "saschac".