Dataverse knowledge in a Prompt action returns one row as 3 separate JSON objects

(1) Share

Report

Posted on by HB-09041311-0

Hi all,
In a Copilot Studio Prompt action (GPT-4.1 mini) I'm using a Dataverse table as the knowledge source. When I check the "Knowledge used" output, I notice that a single Dataverse row is returned as ~3 separate JSON objects that all share the same tls_tlssummaryid, with the columns split across those objects instead of appearing in one record (see screenshot). So ~30 rows show up as ~90 objects, and fields I need together (e.g. summarytext and contenttype) can land in different objects.
To get correct answers and citations, I currently add an instruction telling the model to merge all objects sharing the same tls_tlssummaryid into one record before answering.
My questions:
If merging in the prompt is the only way, is that the recommended approach, or it is not required to specify it?
Is this batched/fragmented retrieval expected behavior, and what causes it?
Is there a setting to return one object per row, or do I always have to merge in the prompt?
Would appreciate any expert guidance.
Screenshot of the "Knowledge used" payload attached.
Thanks in advance!

Categories:

Topic creation & management

General topics

Building Copilot Studio chatbots in Microsoft Teams

I have the same question (0)

All responses (3)

Answers (0)

Sort by

Suggested answer

Mahesh Chintha 168 on at

Like
a
(1)

Report
Copy link

Link copied!

Yes, this is expected. Copilot Studio often retrieves and chunks Dataverse content for grounding, so a single row may appear as multiple objects with the same record ID. As far as I know, there isn't a setting to force a 1:1 row-to-object return. If you need fields from the same record to stay together for accurate reasoning/citations, instructing the model to merge objects by tls_tlssummaryid is a reasonable approach. We follow the deduplicate approach as part of post validation.

Was this reply helpful? Yes No
Suggested answer

Ashlesha-MSFT Microsoft Employee on at

Like
a
(1)

Report
Copy link

Link copied!

Hi, thanks for sharing such a clear repro. What you are seeing is understandable.

Based on current Microsoft documentation and typical Copilot Studio retrieval behavior, this looks consistent with chunked retrieval for knowledge grounding, not a strict one-row-to-one-object return contract for Prompt action knowledge payloads.

Suggested answer to your questions:

1. Is merge logic in the prompt required?
It is not formally documented as mandatory, but your approach to merge objects by the shared row key before answering is a valid and practical workaround.

2. Is fragmented retrieval expected?
Yes, this can happen in retrieval-augmented grounding scenarios, where knowledge is retrieved as chunks rather than guaranteed full rows.

3. Is there a setting to force one object per Dataverse row?
I am not aware of an official Copilot Studio setting that guarantees that output shape in Prompt action Knowledge used.

What I recommend:

1. Keep the merge-by-id instruction in the prompt.
2. Add a rule for missing fields after merge, for example: if summarytext or contenttype is still missing, state that explicitly.
3. If possible, reduce very wide row payloads to lower fragmentation risk.
4. Raise a support ticket with your sample payload if you need Microsoft confirmation whether this is by design for your tenant/model path.

Was this reply helpful? Yes No
Suggested answer

Valantis 6,939 on at

Like
a
(1)

Report
Copy link

Link copied!

Hi @HB-09041311-0,

Your merge-by-id instruction is the right approach.
One optimization to reduce fragmentation: in Dataverse, create a calculated or formula column that concatenates the fields you need together (e.g. contenttype + " | " + summarytext) into a single column. When retrieval chunks the row, the fields that logically belong together travel in the same chunk because they're stored as one column value rather than separate columns.

This is the most reliable way to reduce split-field situations without relying on the model to merge them correctly every time. Fewer wide columns = fewer chunk splits.

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

Was this reply helpful? Yes No