I’m exploring a use case where a UiPath Agent analyzes BRD/requirement documents and generates test scenarios from them.
It works reasonably well with smaller and well-structured requirements, but large BRDs can contain information across acceptance criteria, tables, screenshots, business rules, and different sections of the document.
In these cases, I’m concerned about two problems:
The agent may generate a scenario that sounds valid but isn’t actually supported by the requirement.
It may miss an important requirement because the relevant information is located somewhere else in the document.
For people working with similar Agentic Testing or document-grounded use cases, how are you handling this?
Are you relying mainly on grounding/RAG, splitting the document into requirement-level chunks, adding source references to every generated scenario, or using a separate validation step to verify each scenario against the original requirement?
I’m particularly interested in approaches that work reliably with large real-world BRDs rather than smaller sample documents.
Hi @Supriya_Borah , Chunk the BRD → ground the Agent with relevant documents → generate scenarios → maintain requirement traceability → validate the generated tests. This reduces both unsupported test scenarios and missed requirements.
@Supriya_Borah I’ve also faced similar issues when working with large requirement documents. What helped was not asking the agent to generate scenarios directly from the full BRD.
I first extract the requirements, acceptance criteria, business rules, dependencies, etc., and then generate scenarios from that structured information.
For each scenario, I also try to keep a reference to the actual requirement or section it came from. If the agent cannot find proper supporting information, it should flag it as an assumption or gap instead of creating a scenario.
Another useful check is validating both sides — whether every important requirement has a scenario, and whether every generated scenario actually has a supporting requirement.
For large BRDs, retrieving related sections together also helps, because sometimes the acceptance criteria is in one place and the actual business rule is mentioned somewhere else.
So for me, the biggest improvement came from adding traceability + validation + a human review step, rather than depending only on the prompt.