Q1. What file format should the BI team use?
Use PDF or DOCX — not CSV or XLSX.
Copilot Studio's knowledge index is built for natural language retrieval, not spreadsheet parsing. CSV/XLSX can be indexed but the agent often mismatches column headers with values across many rows.
Best practice: Ask your BI team to export data as a well-structured PDF or DOCX with clear table headers. Even better — convert rows into short prose summaries per record or category for much higher retrieval accuracy.
Q2. Are there size or row limitations?
Yes. Key limits to know:
- Max file size: 512 MB per file
- Practical limit: Large flat files with thousands of rows dilute retrieval quality significantly
- Recommendation: Split files by category, region, or entity type — smaller focused files always perform better
Q3. How should files be structured?
Follow these guidelines for best results:
- Use a clear, consistent header row on every file
- Add a short title/summary at the top describing what the file contains
- Group related data together (e.g., one file per table or category)
- Avoid merged cells — they confuse the parser
- Avoid one massive flat dump of all tables combined
Q4. Does the agent actually perform well on tabular data?
Honest answer: partially.
- General Q&A about the data → Works well
- Specific lookups (find record X) → Inconsistent
- Aggregations / filters (value > X) → Not reliable
Recommended hybrid approach:
- Use SharePoint as a knowledge source for general context and descriptive questions
- Use a Power Automate flow as an agent action/tool for precise, real-time data lookups
