Native way to filter a connector's data per agent?

Forge
, , ,
1

Hi all,

we’ve built a custom Teamwork Graph connector that ingests documents into TWG, and we’re using that data in Rovo agents. So far it’s working nicely.

Now we’d like to filter the search a bit per agent (or subagent), and we’re trying to stick to native approaches if any exist.

The constraints we’ve run into:

  • you can’t have multiple connections of the same custom connector on one site. So we can’t, for example, set up one connector per folder and point each at different folders.
  • and because of that, we can’t give each subagent its own knowledge source (one folder = one source).

So our remaining idea: during sync, we already set a parentKey on each document = the folder it came from. Could we then just prompt the agent to only return/search documents whose parent is a specific folder ID?

Would that even work reliably? i.e. is the parentKey / parent relationship something Rovo search grounding can actually filter on, or is “only use docs with parent X” just a soft instruction the model may ignore?

we’re coming from Querying data from custom connectors - the mechanism already exists (our parentKey creates a parent→child relationship you can traverse in Cypher), but it’s EAP so we can’t use it (connector + API can’t coexist in prod). Since that relationship is there under the hood, can agent prompting hit it - or is grounding completely separate?

Thanks!

2

The parentKey based prompting is indeed a soft instruction that won’t provide hard filtering and may not be reliable as data could still leak reason being the filtering instruction is being executed by the LLM.

3

Thank you for the clarification.

Are there any recommended patterns or upcoming capabilities for restricting a Rovo agent to a specific subset of content from a custom connector? Any “workarounds”?

4

Will surely update with any upcoming capability that would also act as a workaround for this use case. As of now no reliable workarounds exist.