Summary of knowledge sources for agents in Copilot Studio

These knowledge resources enable your agents to provide relevant information and insights to clients.

Published agents contain knowledge that uses pre-configured knowledge sources as the basis for the published agent. You can integrate knowledge at the agent level, in the Knowledge page , or at the topic level.

You can integrate knowledge sources into the agent during the initial creation process, add them after the agent is created, or add them to a Generative Answers node.

Adding and managing knowledge for generated answers

Generated responses allow your agent to find and present information from multiple sources, internal or external, without having to create specific topics. Use generated responses as your primary information source or as a backup when user-generated topics fail to answer a user's query. As a result, you can quickly create and deploy a working agent. Creators don't need to manually create numerous topics, which might not address all customer questions.

By default, when you create an agent, Copilot Studio automatically creates a Conversational Boosting system theme. This theme contains a Generative Answers node, which you can use to start using knowledge sources immediately. All knowledge sources you add at the agent level will be added to the Generative Answers node in the Conversational Boosting system theme .

Supported knowledge sources

Name	Source	Describe	Number of inputs supported in the generated answer	Authentication
Public website	Outside	When a query is entered into Bing, it only returns results from the websites provided.	Generative mode: 25 websites Classic mode: 4 public URLs (e.g., microsoft.com )	Are not
Documents	Inside	Search for documents uploaded to the Dataverse, returning results based on document content.	Generative Mode: All documents Classic Mode: Limited by Dataverse file storage capacity	Are not
SharePoint	Inside	Connect to the SharePoint URL and use GraphSearch to return the results.	Generative Mode: 25 URLs Classic Mode: 4 URLs per topic node Generative Answers	Verify the Microsoft Entra ID of the agent user.
Dataseat	Inside	Connect to the configured Dataverse environment and use enhanced dataverse generation techniques to return the results.	Generative Mode: Unlimited Classic Mode: Two Dataverse knowledge sources (and up to 15 tables per knowledge source)	Verify the Microsoft Entra ID of the agent user.
Enterprise data using connectors	Inside	Connect to connectors where your organization's data is indexed by Microsoft Search.	Generative Mode: Unlimited Classic Mode: 2 for each customizable agent	Verify the Microsoft Entra ID of the agent user.

Note :

1. User authentication for knowledge sources means that when a specific user asks a question to the agent, the agent only displays content that is accessible to that specific user.
2. Knowledge sources within Generative Answers nodes do not currently support Bing Custom Search, Azure OpenAI, or Custom Data. Instead, from the Generative Answers node properties, use the Classic data option for Bing Custom Search, Azure OpenAI, or Custom Data sources.
3. For websites, you need to confirm which websites your organization owns that Bing will search through Copilot Studio.
4. You can perform language-independent queries on all supported file types and languages.
5. If you are using unstructured data, such as individual SharePoint files and folders, OneDrive files and folders, or connectors, there will be different limitations and restrictions. 
6. Currently, citations returned from a knowledge source cannot be used as input for other tools or actions.

Search for knowledge in Classic and Generative modes.

The way the system searches for knowledge sources depends on the mode of coordination the agent uses: Classic or Generative.

Classical arrangement

When you configure an agent to use classic dispatching, the following conditions will apply:

1. In the Conversational Boosting system topic , the number of knowledge sources an agent can search is limited. It depends on the type of knowledge source. Your agent can search any combination of knowledge sources, up to the maximum number specified for each type in the following table:

Type of knowledge source	Limit
Connect to Azure OpenAI Service	5
Bing Custom Search Custom Configuration ID	2
Custom data source	3
Dataverse Knowledge Source	2 data sources, each with a maximum of 15 tables.
SharePoint URL	4
The files have been uploaded.	Unlimited
Website URL	4

1. You can also embed a Generative Answers node into a topic to perform searches for specific intents, rather than just fallback options. The aforementioned knowledge source limitations still apply.
2. Classical orchestration supports custom data sources, in addition to other knowledge sources.

Generating mechanism

When you configure an agent to use orchestration generation, the following conditions will apply:

1. If there are more than 25 different knowledge sources, the agent will filter the knowledge sources using an internal GPT model based on the descriptions provided for each source. 

Note : Files uploaded to the agent are not within the limit of searching 25 knowledge sources.

1. The generative orchestration does not support custom data or Bing Custom Search as knowledge sources. To use those knowledge sources, you must embed them inside a Generative Answers node within a theme.

Enable web search for agents. 

You can find the Use information from the web setting on the Generative AI settings page . You can also find the Web Search setting in the Knowledge section of the agent's Overview page . This setting allows your agent to access broad, real-time, and up-to-date information beyond what's available in predefined or enterprise-specific knowledge bases. This setting requires the agent to have orchestration enabled.

When you enable the "Use information from the web / Web Search" setting , it will be triggered when a user's question could benefit from information on the web. It will search all public websites indexed by Bing. This type of search occurs in parallel with any searches on public websites you've added as knowledge sources. Results from " Use information from the web / Web Search" are interleaved with results from your configured public website knowledge sources.

Note : Use information from the web / Web Search  uses Grounding with Bing Search to return information from the web.

Allow unsubstantiated feedback.

Setting "Allow ungrounded responses" in the Knowledge section of the agent's Generative AI settings manages whether the agent can generate responses using only the model's general knowledge. This setting requires the agent to have generative facilitation enabled.

When you enable the Allow ungrounded responses setting , your agent can respond using the model's general knowledge, even if it doesn't use any knowledge sources or tools.

When you disable this setting, the agent will block any responses generated within a conversation that it doesn't use a knowledge source or tool for. This means that if the agent decides to answer a question directly from its conversation history or general knowledge, without calling a knowledge source or tool, the response will be blocked and the fallback thread will be triggered.

For example, consider the following conversation:

1. User: "What is the return policy for online orders?"
2. Agent: "Our return policy allows returns within 30 days of purchase (for all items, including discounted items)." (Taken from knowledge source)
3. User: "Does that apply to discounted items as well?"
4. Agent: (Blocked)

In this case, the agent might decide that it already has enough context from previous turns to answer the next question without calling any additional tools or knowledge sources, so the response is blocked.

Note : Disabling the "Allow ungrounded responses" setting does not guarantee that the agent will never use common knowledge. The agent's underlying model may still incorporate common knowledge when it combines this knowledge with information it gathers from knowledge sources or tools. This setting only blocks responses where the agent does not use any knowledge source or tool in that turn.

Turn on Work IQ

The "Turn on Work IQ" setting on the Generative AI settings page determines whether your agent uses semantic search to improve search results. This setting requires the agent to have dispatch generation enabled.

This feature requires the agent to share a tenant with the Microsoft 365 Copilot license. It also requires semantic indexing to be configured for use. To use semantic indexing, the Microsoft 365 Copilot license must be assigned to at least one user within the enterprise.

Important note : The Turn on Work IQ feature requires the agent's user authentication to be set to "Authenticate with Microsoft" . If authentication is set to any method other than "Authenticate with Microsoft" , you cannot change this setting.

When you enable this feature and the creator has a Microsoft 365 license in the same tenant, the agent will support SharePoint and connectors containing files up to 200 MB in size. This feature is enabled by default.

Note :

1. For SharePoint knowledge-based agents, the Turn on Work IQ feature provides significantly improved knowledge retrieval and response quality. This feature utilizes advanced internal retrieval tools that allow agents to gather a greater amount of context with higher accuracy. However, due to the increased system complexity, some users and queries may experience a slight increase in latency.
2. If you don't have a Microsoft 365 Copilot license in the same tenant as your agent, or you're experiencing lower response quality, turn this feature off.
3. Agent creators do not need a Microsoft 365 Copilot license to create agents with semantic indexing.
4. SharePoint and Microsoft Copilot connectors support files up to 512 MB in size if they have the extensions PDF, PPTX, or DOCX. 
5. The Turn on Work IQ feature is a separate feature from the Dataverse search feature. 

Source verification

If you are using SharePoint, Dataverse, or enterprise data with Microsoft Copilot connectors, you need to integrate authentication. 

Additionally, you may need to consider URL elements that require extra validation for your source.

Content censorship

Content moderation settings allow your agent to provide more responses. However, increasing the number of responses may affect the agent's ability to allow malicious content.

You can configure content moderation settings in the following areas:

1. In the Generative AI settings page, configure moderation at the agent level.
2. The settings in the Generative Answers node configure moderation at the topic level.
3. The settings in the prompt tool set the moderation mode to the prompt level.

During runtime, theme-level settings will take precedence. If you don't set theme-level content moderation, it will default to the Generative AI settings .

To override the content moderation settings of an agent or topic for prompt tools, configure the prompt tool's Completion setting to send a specific response.

To adjust content moderation settings at the agent level, change the agent's Generative AI option to Generative .

1. Choose your desired level of moderation for your agent.

Content moderation levels range from Lowest to Highest . The lowest level generates the most responses, but these may contain harmful content. The highest content moderation level generates fewer responses and applies stricter filters to limit harmful content. The default moderation level is High .

2. Select Save .

To adjust content moderation settings at the topic level, change the settings in your Generative Answers node.

To adjust content moderation settings for the prompt tool, change the settings in the prompt builder.

Official sources

When adding knowledge sources to your agent, you may not always have control over how the information evolves over time, or you may not fully trust this information. It's important to let your users know that they should consider the answers with caution and that they should verify them when appropriate.

However, when you know that information from a particular knowledge source has undergone a rigorous verification process and is highly reliable, you can mark that source as official and use it directly without further verification.

To mark a knowledge source as official, on the Knowledge page , select the three dots (⋮) for the knowledge source, hover your mouse over Official source , and select Yes .

Note :

1. This feature is not yet compatible with orchestration. If you want your agent to use official knowledge sources and mark them as such, disable orchestration.
2. When an agent utilizes authoritative knowledge sources, the response will begin with a distinctive signal.