Power Platform Community Forum Thread Details

I'm sure it worked fine before, but the prompt tool suddenly malfunctioned today
No matter what filtering conditions I input, it always first retrieves a fixed thirty records from the dataverse and then filters them based on my input criteria
And this definitely cannot be used

As you can see, no matter which column or keyword I search for, the knowledge it refers to is the same, which naturally makes it impossible to find the correct answer
My expected outcome is to first get the 30 most relevant ones, and then query the answers
The current result: Regardless, take 30 first and then query inside
Does anyone have any ideas? Thank you

Categories:

Bot extensibility

Hello,
This behavior isn’t due to any misconfiguration on your part but a current limitation/regression in the Copilot Studio Prompt Tool when used with Dataverse knowledge: instead of applying filters and relevance ranking in Dataverse, the tool always retrieves the same first 30 rows (in an internal order) and only applies filters in memory afterward.
As a result, the “knowledge used” never changes, different queries reference the same records, and correct answers are impossible if the relevant data isn’t in those initial 30 rows—making changes to columns, keywords, prompts, or models ineffective. This surfaced with the Prompt Tool + Code-gen preview rollout, which hard-caps retrieval at 30 records and does not push $filter, $search, or $orderby server-side, with no available settings or prompt-level workaround. The practical, production-safe approach is to move retrieval out of the Prompt Tool—use Power Automate or the Dataverse/Web API (or Azure Cognitive Search for true relevance) to apply server-side filtering and ranking, then pass the correctly scoped results to Copilot for reasoning.