How to enforce a custom Rovo agent to search for TECH Document information for a legacy job?

[Narmatha A]

I have created a Custom ROVO agent . Purpose of the agent is if tidal job is an input means response should return all the informatica components if exist and suc/TDD documents confluence link .

scenario1:
If TDD link doesn’t exist for a legacy job it should check all jira stories does anything have this tidal job mentioned and return the link . General Agent is able to track it but my custom agent not doing it.

scenario 2:

There is another scenario even everything all exist . It is returning java TDD,Azure TDD links not associated with my tidal job or infa component. 

Scenario 3 : Similar to Scenario 2 tried different job it had all details accurate but returned different job informatica components 

[Viswanathan Ramachandran · Accepted solution]

Hi @Narmatha A 

For scenario 1

Root Cause: Custom agents do not perform fallback multi-step searches unless explicitly commanded in the system instructions. 
Solution: Define an explicit Fallback Execution Chain and explicitly enable your search tool in your Agent settings.

For scenario 2

Root Cause: Rovo uses semantic similarity search. If a direct exact match for your Tidal job is weak, the vector search returns generic technical documents (like Java/Azure TDDs) because they contain similar words (e.g., "TDD", "Job", "Pipeline").
Solution: Enforce Exact String Matching and strict Relevance Filtering in the instructions.

For scenario 3

Root Cause: The agent is retrieving the correct details but its extraction context window is pulling Informatica component names from neighbouring tables, templates, or related sections within the same space.
Solution: Bound the Data Extraction Scope within your prompt instructions.

This checklist below might help

Field	Recommended Action
Knowledge	Attach specific space/project where Tidal/Informatica details live.
Tools	Ensure both Search tools are attached.
Instructions	Implement strict string matching, fallback execution steps, and extraction boundary rules.

[Narmatha A]

Hi @Viswanathan Ramachandran I need one more suggestion. In my custom agent is facing some tech escalation failure if TDD exists for a tidal job but it prefers SUC as primary but in that page TECH clickable link also exists in the page . Nearly tested 5 jobs behavior same

[Viswanathan Ramachandran]

Hi @Narmatha A 

This issue happens because Rovo’s retrieval model prioritises the first overarching document structure it encounters (in your case, the SUC page) and treats embedded hyperlinks inside the content as passive text rather than active navigation targets.
To fix this, you need to explicitly instruct the agent on how to handle Page Hierarchy and Link Following when a SUC page is found.

Update your agent's Instructions to include explicit logic for detecting and opening embedded Technical Links within SUC pages.

In summary,

Explicit Tool Execution: Rovo will not automatically trigger a secondary API call to fetch page contents linked inside a retrieved page unless explicitly told to treat embedded URLs as actionable targets.

Overrides Structural Bias: Rovo naturally prefers top-level parent pages (SUCs) because they often match job names directly in the page title. The directive forces it to look deeper into the page body.

[Narmatha A]

Hi @Viswanathan Ramachandran  , ROVO created in my client account. Even after implementing rules it’s not working. I am facing inconsistency and behavioral issues with my agent. If a confluence page is not maintained properly it’s messing up.

[Narmatha A]

@Viswanathan Ramachandran  if I share my instructions and expectations you will be able to help me to reframe my instructions 

[Dr Valeri Colon _Connect Centric_]

@Narmatha A welcome to the community. I’d treat this as a retrieval/knowledge-quality issue rather than assume a specific Rovo ranking behavior. Narrow the agent to the exact Jira projects/Confluence spaces it should use, require exact matching on the Tidal job/component name, and instruct it to return “Not found” rather than substitute similar documents. If General Chat finds the right content but the custom agent doesn’t, compare the agent’s Knowledge/Tools and use Debug Response to see what sources it actually retrieved.