← Back to Selected Work

Building a Production RAG Support Assistant

~30% fewer support tickets Onboarding satisfaction ↑ 37% Continuous learning loop established

Executive Summary

Helped define and ship a production RAG-based support assistant for Oracle Guided Learning and Oracle Cloud Success Navigator, grounding responses in enterprise content with confidence gating and human escalation, reducing support tickets by approximately 30% across a 15M+ user ecosystem.

15M+ describes the wider Oracle Guided Learning and Cloud Success Navigator ecosystem the assistant served, not a verified count of direct assistant interactions.

Business Problem

Complex Oracle Fusion workflows generated high support ticket volume, and onboarding friction was reducing platform engagement across enterprise clients. Support content existed but wasn't reaching customers at the moment they needed it.

Why It Mattered

Every unresolved question that became a support ticket cost time on both sides, and slow onboarding meant customers were reaching value later than they should.

My Role and Scope

Helped define and ship the assistant as part of the broader Product Operations & Strategy role at Oracle, working with engineering on retrieval architecture and with the content organization on GenAI-assisted content creation.

Constraints

Answers had to be grounded in approved enterprise content, not generated freely, given the accuracy and trust requirements of an enterprise support surface. Coverage had to hold up across complex, frequently changing Fusion workflows.

Decisions I Made

  • Chose a retrieval-augmented approach over open generation so every answer traced back to approved content.
  • Built confidence gating so the assistant would escalate to a human rather than guess when it wasn't sure.
  • Paired the assistant with GenAI-assisted content creation so teams could scale contextual in-application guidance without a proportional increase in content-authoring effort.

How the Product Worked

The assistant retrieved relevant enterprise content via vector search, applied confidence gating to decide whether it had a grounded answer, and escalated to a human when it didn't. GenAI-assisted content workflows kept the underlying content current for complex Fusion workflows while maintaining quality and governance controls.

Cross-Functional Leadership

Partnered with engineering on retrieval and confidence-gating architecture, and with the content and support organizations to define escalation paths and content governance.

Outcome and Evidence

  • Reduced support tickets by approximately 30% across the 15M+ user ecosystem.
  • Onboarding satisfaction improved 37%.
  • Established a continuous improvement loop using behavioral data to refine the assistant over time.

What I Learned

A grounded, willing-to-escalate assistant earns more trust than one that always has an answer — confidence gating was as important to adoption as retrieval quality itself.

Related Capabilities

  • RAG
  • Vector retrieval
  • Confidence gating
  • Human-in-the-loop design
  • Enterprise content strategy

Think this experience maps to a role you're hiring for?