Fixed a production outage caused by a 100 MB connector buffer cap, and turned the corrected design into a reusable template that worked first time.
The problem
The agent worked perfectly on a sandbox with a dozen transcripts. In production it failed outright: the flow pulled every transcript body in one query and blew through the connector’s 100 MB buffer.
What I did
- Re-architected the search flows: the bulk query now fetches IDs only, with a 90-day date clamp and pagination circuit breakers; bodies are fetched per record in a loop capped at five concurrent calls.
- Moved transcript matching to Dataverse Relevance Search, and added flags so the agent tells users honestly when results may be truncated.
- Packaged the proven design as a customer-agnostic template agent with generic instructions and placeholders.
Result
Production search works again, and the template agent returned correct sentiment tables and themes on its very first live test.