Copilot Studio Agents

(2) Share

Report

Posted on by CK-25081024-0

Hello,

I wanted to understand how document ingestion works in a Copilot Studio agent.

When we provide documents as knowledge sources, does the platform internally implement a RAG pipeline (including chunking, embedding generation, indexing and retrieval) or does it rely on a different mechanism?

Thanks in advance for your insights.

-C

Categories:

Autonomous agents

All responses (5)

Answers (3)

Sort by

Verified answer

MS.Ragavendar 7,798 Super User 2026 Season 2 on at

Like
a
(0)

Report
Copy link

Link copied!
@CK-25081024-0,

Document Upload ->Ingestion into Dataverse ->Chunking (split into smaller pieces) -> Embedding Generation -> Vector/Semantic Index Creation -> User Query ----> Semantic Retrieval of Relevant Chunks -> LLM Generates Grounded Response

For uploaded documents and most unstructured knowledge sources, Copilot Studio itself implements a managed RAG pipeline consisting of:

Chunking

Embedding generation

Semantic/vector indexing

Retrieval of relevant chunks

LLM-grounded answer generation

All of this is handled automatically by the platform, so makers generally do not need to build or manage the RAG infrastructure themselves

Do Developer / Agent Creator control chunk size, embedding model, or indexing settings?

For the built-in knowledge experience is No, This has been managed by Microsoft.

Chunking strategy

Embedding generation

Index creation

Retrieval orchestration

These are not exposed as configurable settings in standard Copilot Studio knowledge sources.

More Information :

https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-unstructured-data

https://www.microsoft.com/en-us/power-platform/blog/2025/03/27/knowledge-in-microsoft-copilot-studio/

https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-copilot-studio

✅If this helped, please Accept as Solution to help others ❤️ A Like is appreciated 🏷️ Tag @MS.Ragavendar for follow-ups.
Verified answer

Valantis 7,545 Super User 2026 Season 2 on at

Like
a
(1)

Report
Copy link

Link copied!

Hi @CK-25081024-0,

Adding to the above. For the built-in knowledge sources, chunking and embedding settings aren't configurable. But if you need control over those parameters, you can bypass the built-in RAG pipeline entirely by using Azure AI Search or Azure AI Foundry as a custom knowledge source.

With Azure AI Search: you build and manage your own index with your preferred chunking strategy, embedding model (e.g. text-embedding-3-large), and retrieval settings. You then connect it to Copilot Studio as a knowledge source via the Azure AI Search connector. This gives you full control over chunk size, overlap, semantic ranking configuration, and the embedding model used.

With Azure AI Foundry: you can build a custom RAG pipeline using Prompt Flow or the Foundry SDK, then expose it as an API endpoint or MCP tool that the agent calls. This gives you the most flexibility including custom retrieval logic, hybrid search, and reranking.

So the answer is: no control within the built-in knowledge experience, but full control if you use Azure AI Search or Foundry as the backing infrastructure instead.

Best regards,

Valantis ✅ If this helped solve your issue, please Accept as Solution so others can find it quickly.

❤️ If it didn’t fully solve it but was still useful, please click “Yes” on “Was this reply helpful?” or leave a Like :).

🏷️ For follow-ups @Valantis.

📝 https://valantisond365.com/

💼 LinkedIn ▶️ YouTube
Verified answer

11manish 4,802 Super User 2026 Season 2 on at

Like
a
(0)

Report
Copy link

Link copied!

Copilot Studio provides a managed RAG capability rather than requiring developers to build the complete RAG pipeline themselves. When documents are added as knowledge, Copilot Studio handles the ingestion, processing, chunking, indexing/embedding and retrieval infrastructure. At runtime, the user's query is optimized, relevant knowledge is retrieved, and the retrieved content is supplied to the language model to generate a grounded response. The exact retrieval and indexing implementation depends on the knowledge-source type.
Haque 4,280 Super User 2026 Season 2 on at

Like
a
(0)

Report
Copy link

Link copied!

@CK-25081024-0

Top of everything I have a question - what specific file types or data sources (PDFs, SharePoint, Salesforce, etc.) are you planning to ingest?

There are some differences based on the knowledge source how copilot studio behaves with the given knowlege.
CK-25081024-0 24 on at

Like
a
(0)

Report
Copy link

Link copied!

@Haque We have added PDFs containing both text and images to the agent’s knowledge sources. Does the processing of the content change when images are present in the documents?

And are the RAG parameters fixed across all agents and data sources? For example, the embedding model and dimensions, chunking strategy, number of retrieved passages etc...

We are currently testing the agent as a chatbot integrated into our application and we don't have the best performances that is why we wanted to better understand what happens in the background when a user asks a question.