Answered

Using the Copilot connector to connect Zendesk Guide in a Multibrand environment (no answers)

(1) Share

Report

Posted on by Daniel Winter

Hi,

I'm working with Co-Pilot Studio to integrate our Zendesk Guides, allowing employees to search all of our content directly. Some content is restricted to logged-in users assigned to specific segments, so the bot needs to parse both public and restricted content.

We have multiple Help Centers, and by using the appropriate subdomains, I’ve successfully connected them all. I’ve selected only the "Articles" table for synchronization, as it contains the relevant content. The connection appears to be working, and I can preview the content successfully:

Preview of articles table. Body holds the html content

For testing I have used a question that reflects the title formed as a question to be able to return content from Gen AI.

I see that it's looking at the knowledge, but not returning any results:

These are my Gen AI settings:

Use generative AI to determine how best to respond to users and events (preview): ON
Allow the AI to use its own general knowledge: OFF

How should your agent interact with people?: Generative

How strict should the content moderation be?: HIGH

My expectation was that it's able to parse through the Articles table and providing me at least some content. Any suggestions what I'm doing wrong here?

Categories:

General topics

I have the same question (0)

All responses (9)

Answers (1)

Suggested answer

Artur Stepniak 1,539 Super User 2025 Season 2 on at

Like (0)

Report

Hello,

at first it's worth to note that generative orchestration (this buddy: Use generative AI to determine how best to respond to users and events (preview): ON) is a preview feature - it means that it couldn't work in some cases or give unexpected results, see docs: https://learn.microsoft.com/en-us/microsoft-copilot-studio/advanced-generative-actions.

Second point: did you input any instructions to the model on how to perform tasks? I see that the columns have system names, so it's not able to understand what do you mean. You need to describe it in the instructions on what columns you have, what to do with them etc. If that won't work, you can also try to use Actions instead of a connector. I see Zendesk connector on the list:

If you'd use an action, you'd need to instruct it on how to prepare inputs/outputs.

In case of any other questions, let me know. If the answer helped you, mark it, so that others can benefit from it.

Best regards,

Artur Stepniak

Was this reply helpful? Yes No
Daniel Winter 19 on at

Like (0)

Report
Thanks for your feedback. I did now rewrite my instructions for the model like this to include the concrete column names of the tables that are imported:

You will be supporting our team in answering questions around our product. Your primary knowledge source are answers based on Zendesk Guide articles.

Key Information Sources:

Zendesk Guide Articles: You will primarily use Zendesk Guide articles as your knowledge source. Each article is structured in a table with the following columns:

title: The title of the article.

body: The full content of the article, which may include HTML formatting.

locale: The language in which the article is written.

html_url: The direct URL link to the article.

Answering Guidelines:

Use the Articles: Base your responses on the Zendesk Guide articles available in the table.

Provide Accurate and Concise Answers: Your answers should be both correct and to the point.

Always Include a Source: Whenever you use an article to answer a question, always include the HTML URL as a reference. This ensures transparency and allows others to verify the information.

Example Response:

If the article titled "Feature XYZ Overview" provides the relevant information, your response should be something like:

••"Feature XYZ allows you to... (provide summary of the key points)."

Source: Feature XYZ Overview

This did not help to find content. I also turned Use generative AI to determine how best to respond to users and events (preview) OFF.

It is picking the GenAI topic as fallback where the knowledge content is configured:

Was this reply helpful? Yes No
Suggested answer

Artur Stepniak 1,539 Super User 2025 Season 2 on at

Like (0)

Report

Hello,

I'd experiment with Actions instead of using a connector to knowledge source. You can try then to turn on Generative orchestration. That way the model would fill in the inputs of an action by itself and retrieve the articles, and output the response.

Other way round you could use the classic approach with trigger phrases and the action to filter out the articles based on the user prompt. You can then use generative answers action to only generate a response to the user with the data that was retrieved. This option is a bit safer in terms of stability.

In case of any other questions, let me know. If the answer helped you, mark it, so that others can benefit from it.

Best regards,

Artur Stepniak

Was this reply helpful? Yes No
Daniel Winter 19 on at

Like (0)

Report

Hey,

so I tried using actions with the onboard Zendesk connector where I used the search articles. Sometimes it did find something but this was not always the case.

I went know back to using the native knowledge source, but instead of using the built-in Zendesk connector, I essentially exported all articles per brand to a dedicated json where I just added three attributes: Title, Content, ContentLocation (holds the external URL to the article) .

I uploaded these files now manually for the purpose of this test. Now the agent returns good results during my initial tests. I only have now one problem.

Instead of using the ContentLocation as the reference (citation), it often just references the whole json file.

Within the knowledge source I added some description like this already:

It is structured in a JSON format with the following three key attributes:
* Title: The article’s title, summarizing the key topic or feature.
* ContentLocation: The URL where the article is hosted, to be used as a reference or for direct access to the content.
* Content: The full body of the article, presented with HTML formatting, containing in-depth details and explanations of the feature.

For citation you should always use the ContentLoaction as the reference.

I also adjusted the instructions..but still it doesnt provide the right reference...If i ask for the specific url o the reference it returns it.

So I have two questions:

1. Is it possible to get the agent to actually use the URL as the reference

2. Is there any way on updating the json through an automated way (e.g. power automate flow that just updates the json on a daily/weekly basis)

Was this reply helpful? Yes No
Verified answer

Artur Stepniak 1,539 Super User 2025 Season 2 on at

Like (1)

Report

Hello Daniel,

unfortunately, you won't be able to guide the model to show an exact reference to the URL this way. It's because the file, that you've provided to the model is indexed and the reference is pointing to the indexed content called embeddings. The pointer is based on the user prompt. To achieve better results, you could switch the datasource to a Sharepoint List or Dataverse table. The model would filter out the content row by row by using ODATA filter. It could output the URL stored in the column then.

The second thing is that you won't be able to easily update the JSON file, you'd also need to upload it manually each time it's updated. It's not efficient overall.

In case of any other questions, let me know. If the answer helped you, mark it, so that others can benefit from it.

Best regards,

Artur Stepniak

Was this reply helpful? Yes No
Daniel Winter 19 on at

Like (0)

Report

Just wanted to give feedback on this. I created now a Dataverse table where Title, Content and the URL are stored. I added this as a knowledge source and removed all the other sources. Sadly it did have the same effect as using the native Zendesk connector. Meaning it did not find any results. Any suggestions what might be the problem? On the source status it shows as "Ready"

Was this reply helpful? Yes No
Suggested answer

Artur Stepniak 1,539 Super User 2025 Season 2 on at

Like (0)

Report

Hello,

you'd need to use instructions to guide the model on how to use the datasource. Have you also described it? It needs to have some kind of description to understand what is it. When adding a datasource, you can describe the columns, table itself etc., see:

You'd also need to prompt it correctly as it's very important. Based on descriptions that you prepare, you should prompt the model. Without this it won't understand what columns to use, what table etc.

If it won't work, you can also try to experiment with turning on generative orchestration, it's a preview feature though so you should keep that in mind.

In case of any other questions, let me know. If the answer helped you, mark it, so that others can benefit from it.

Best regards,

Artur Stepniak

Was this reply helpful? Yes No
Daniel Winter 19 on at

Like (0)

Report
I did. Not sure how good they are for the bot though.

Here is what I have:

Instructions

You are an expert on Radancy’s Talent Acquisition Cloud Platform, a SaaS solution in the HR space. Your task is to answer customer questions during RFPs, focusing on all aspects of the platform. The platform comprises multiple features designed to optimize talent acquisition, including:

Employee Referrals

Hiring Events

Career Site

CRM

Insights & Analytics

Programmatic AdTech

Key Information Sources

Your answers should be based on articles provided in a Dataverse table. Each row includes:

Title: Name summarizing the content.

Body: Full HTML-formatted content.

Feature: Related sub-product/feature.

HTML URL: Direct link to the article (always share this).

URL: Internal URL (do not share).

Zendesk Article ID: Internal identifier.

Locale: Language of the article.

Answering Guidelines

Use Articles as Sources: Always derive your responses directly from the provided articles.

Be Accurate and Concise: Provide clear and precise answers without unnecessary detail.

Always Include the Source: Cite the HTML URL of the referenced article to maintain transparency and verification.

Example Answer Structure

Use the following structure when responding:

<ANSWER TO THE QUESTION>

You can find more information in this article: [HTML URL]

Example

Question: What features does the CRM module offer?

Answer: The CRM module provides tools for candidate engagement, talent pipeline management, and automated workflows to enhance recruiter efficiency.

You can find more information in this article: [https://example.com/crm-article]

Remember, always cite the HTML URL and ensure your answers are derived directly from the articles provided.

Within the Dataverse Source I have added this description:

The Dataverse table serves as the knowledge source for all published Zendesk articles related to the Talent Acquisition Cloud Platform. Each article corresponds to one of the platform's features, including:

Employee Referrals, Hiring Events, Career Site, CRM, Insights & Analytics, Programmatic AdTech

Key Columns in the Table:

Feature: Identifies the platform feature associated with the article.

Title: The article's title, summarizing its content.

Body: Contains the full content of the article, formatted in HTML.

HTML URL: Direct link to the full article (for external sharing).

URL: Internal system link (should not be shared).

Zendesk Article ID: Internal ID for tracking the article in Zendesk.

Locale: The language of the article.

The Body column contains the core content of each article, which should be referenced when crafting responses. Always cite the HTML URL to maintain transparency and allow access to the full article.

As an addition I added descriptions to the columns as well.

Was this reply helpful? Yes No
CU26110138-0 16 on at

Like (0)

Report

Looks like Copilot is struggling to match your query to the HTML it indexed, which usually happens when the content is still too markup heavy. I work at eesel AI and we see the same thing sometimes when Zendesk Guide articles aren’t normalized to plain text. Stripping the body down to clean text or creating a simplified view in Zendesk usually fixes the retrieval.

You could also try loosening moderation for a minute just to confirm the model can return something before tightening it again. If the preview looks good, it’s almost always an embedding mismatch rather than a permissions issue.

Was this reply helpful? Yes No