Answered

Dataverse Retrieval Issue: New Records Not Used by Copilot Studio Agent

(2) Share

Report

Posted on by HB-09041311-0

Hi Community,

I am observing some unexpected retrieval behavior in Copilot Studio agent and would like to validate whether this is related to Dataverse Search or something else in the retrieval pipeline.

Scenario-

I have a Dataverse table containing knowledge from multiple sources.
About 1,000+ Confluence records were ingested 2 days ago.
About 500+ Jira records were ingested today into the same table.
My agent uses an AI Builder Prompt node, where Dataverse table columns are passed into the prompt.

Observation

When I ask questions that should be answerable from the Jira content, the agent always consistently responds using information from the Confluence records. It appears that the newly ingested Jira records are never being selected.

Questions

If Dataverse Search is disabled, could that be the primary reason Jira records are not being retrieved, even though they exist in the table?
Does enabling Dataverse Search allow Copilot Studio/AI Builder scenarios to leverage semantic retrieval and relevance-based ranking, rather than relying on basic filtering or keyword matching?
For mixed-source datasets (Confluence + Jira), are there any known retrieval limits, ranking behaviors, or indexing considerations that could cause one source to consistently dominate the results?
Has anyone observed similar behavior where recently ingested records are ignored until Dataverse Search is enabled or fully indexed?

At this point, I'm trying to determine whether this issue is due to:

Disabled Dataverse Search,
Incomplete indexing of newly ingested records,
Or a retrieval/ranking limitation within the AI Builder Prompt + Dataverse approach.

Any insights or validation from others who have implemented similar architectures would be greatly appreciated.

Thanks in advance!

Categories:

General topics

Calling actions from Copilot Studio

All responses (3)

Answers (1)

Sort by

Verified answer

Dua Sadiq 129 on at

Like
a
(2)

Report
Copy link

Link copied!
Hi,

These are all documented behaviors in Microsoft's Power Platform docs to better understand and see if it resolve your queries.

Dataverse Search being disabled is likely the primary cause. Microsoft explicitly states: "All generative AI-powered experiences, such as Copilot and Agents, require Dataverse Search." Without it, there's no semantic ranking records are pulled in basic order, which means whichever source was ingested first wins consistently. That explains why Confluence always dominates.

📄 Reference: https://learn.microsoft.com/en-us/power-platform/release-plan/2025wave1/data-platform/enhance-ai-powered-experiences-dataverse-search

The 30-record default limit is also working against you. Per AI Builder docs, the prompt node only fetches 30 records by default. With 1,000+ Confluence records ingested before your Jira records, the first 30 pulled are almost certainly all Confluence. Microsoft's recommendation for tables over 1,000 records is to use FetchXML to filter down to the relevant subset before the prompt sees the data.

📄 Reference: https://dev.to/balagmadhu/querying-dataverse-using-ai-builders-grounded-prompts-1oco

Indexing delay is the third factor. Microsoft docs note indexing can take anywhere from 30 minutes to several hours for large bulk ingestions so your Jira records ingested today may simply not be in the search index yet.

📄 Reference: https://learn.microsoft.com/en-us/microsoft-copilot-studio/knowledge-unstructured-data (under limits and limitations)

Recommended fix:

Enable Dataverse Search in PPAC if not already on

Wait for Jira records to finish indexing

Add a FetchXML filter in your prompt node scoping by source (source = "Jira")

Hope it helps if it does i would appreciate it if you mark it as verified answer.
Suggested answer

Mohsin Ali 1,075 on at

Like
a
(1)

Report
Copy link

Link copied!

Hello @HB-09041311-0 - the dataverse search can be first thing that I would validate. However, since your older confluence records are being returned while the recently added Jira records are not, I would recommend to turn off/on the search once, publish the table and see the impact in your test pane before.

Also, can you please confirm if there was recent customizations were made. If so then that relevant information would not come up as response from the Copilot Agent.
Mohsin Ali 1,075 on at

Like
a
(0)

Report
Copy link

Link copied!

Hello @HB-09041311-0 - Just wanted to follow-up with you if the suggestion I provided helped you or in case you have any further questions.

Otherwise, please do me a favor and mark my suggestions as verified. This will help the community to navigate quickly.