There are two separate problems here and it's worth addressing both.
Problem 1: Excel is not a reliable knowledge source
This is almost certainly why your knowledge isn't being retrieved. Microsoft docs and community testing both confirm that Excel files index significantly less reliably than Word or PDF in Copilot Studio. The same tabular content placed in a DOCX or PDF will index correctly where the XLSX version is skipped entirely, even when both files are in the same SharePoint folder. Microsoft explicitly notes that Excel is supported but responses to structured data queries may not be optimal.
Convert your QnA files to Word (.docx) or PDF and re-add them as knowledge. That alone is likely to fix the retrieval issue.
Problem 2: Knowledge is the wrong place for clarification logic
Even if the Excel files were indexed correctly, using a QnA row to return "Which country would you like information for?" is not how this should work. Knowledge sources are designed to return answers, not to drive conversational branching. The agent retrieved the answer from reasoning (you can see it in the internal trace) but the generative model then couldn't use it as a final response because it's not structured as an answer to the user's question, it's a question itself.
The clarification logic should live in a Topic, not in the knowledge base:
- Create a topic triggered by phrases like "benefits", "what are my benefits"
- Add a Question node: "Which country would you like information for?"
- Based on the answer, call a Generative Answers node with a more specific query like "benefits [country]" against your knowledge source
This gives you the clarification flow reliably without depending on the knowledge retrieval to return a question.
Your questions:
1. Re-indexing frequency: automatic, asynchronous, no confirmed fixed schedule. Can take minutes to hours. No manual trigger available.
2. Force refresh: not possible. Removing and re-adding the knowledge source is the closest thing to forcing a re-index.
3. Can fallback override correct logic: yes. If the knowledge retrieval returns no usable content, the fallback fires regardless of what was in the internal reasoning trace.
4. Excel format requirements: Excel is supported but not first-class. Word or PDF is significantly more reliable for retrieval.
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/
