Power Platform Community Forum Thread Details

I am creating an agent and have noticed a few issues happening whenever an answer is generated from a topic:

1. I sometimes get: Error Message: The content was filtered due to Responsible Al restrictions. Error Code: ContentFiltered Conversation

- I have web search turned off for the topic (and agent)

- I have set the topic to ONLY search through my knowledge sources which is all technical documentation

- Only setting the content moderation level to "low" gets rid of it, but sometimes it still shows up

2. As of today I will also get back unformatted walls of text:

Ex: Creating a Button Step-by-Step Instructions: 1. Understand Button Anatomy and Usage - Here is some information about the button. 2. Choose the Variant - here is some information about the vaariant

instead of what I used to get:
Creating a button step by step instrucitons:

1. Understand...

2. Choose the variant....

3. Sometimes I will get a log after an answer:

{"new_instruction":"Select a system your question relates most to","explanation_of_tool_call":"This action needs to be done because the tool execution message introduced an instruction to prompt the user to select a system, which was not explicitly requested by the user."}

I don't have any of these issues if the agent responds and does not hit a topic that responds with a generative answer.

Categories:

Topic creation & management

Hi @ DC-12021822-0

This sounds like a mix of how generative answers in topics are currently handled and some platform quirks. The content filtering can still trigger even without web search if the docs contain security-sensitive terms or very procedural instructions, since safety filters apply to all sources. The formatting issue is often prompt-related — it helps to be very explicit about the role, process, and output format (for example: “You are a technical documentation assistant. Summarize clearly using markdown, headings, and numbered steps, each on a new line”). The JSON showing up is likely internal tool/debug output leaking into the response; adding a final instruction like “return only the final user-facing answer, no JSON or system messages” can reduce that. I’d first test by tightening and clarifying the prompt, and if it still happens, it’s probably a backend issue worth raising with support along with a sample run.