Copilot Studio Agent – Poor Results with CSV Knowledge Source (Summarization & Categorization)

Hi team,
I'm building an agent in Copilot Studio using a CSV file (~40MB, JIRA extract) as the knowledge source. The goal is to enable the agent to summarize, categorize, and extract insights from the data for specific areas.

To improve performance, I’ve also tried converting the CSV into smaller PDFs. However, the results remain unsatisfactory—especially for user queries involving summarization and categorization.

Interestingly, when I compare this with Copilot Analyst/Researcher, the responses are significantly better and more relevant.

I’m wondering:

Are there any fine-tuning options, configuration settings, or best practices we should consider to improve Copilot Studio agent performance?
Is there a recommended approach for handling large structured data like CSVs in Copilot Studio?
Would appreciate any guidance or suggestions to help move forward.

Thanks in advance!

The issue names a roughly 40MB JIRA CSV and smaller PDFs used as Copilot Studio knowledge sources, but mentions no repository file, test, or entry point. Start by reviewing the agent's knowledge-source configuration and compare CSV and PDF query behavior; completion is undefined until the requested guidance or configuration target is agreed.