AI Documentation Assistant

RBC · Senior Product Manager

AI Documentation Assistant

512 → 37

Repositories → Wikis

~40% faster

Engineer Onboarding

~30% fewer

Support Questions

AI AgentsKnowledge ManagementDeveloper ExperienceRAGResponsible AI

Overview

Built repository-reading agents that produced evidence-backed service wikis, then proposed a citation-first knowledge assistant for employees based on those structured sources.

The Challenge

Platform knowledge lived in people’s heads and undocumented systems, slowing onboarding and incident response. Documentation was costly to maintain, and a useful assistant needed evidence-backed answers rather than unsupported prose.

The Approach

Designed agents to generate standardised wikis with evidence, ownership, confidence and review metadata, updating only pages affected by repository changes. Proposed a retrieval assistant with citations on every answer, an explicit “I don’t know,” expert review below a confidence threshold and a read-only pilot.

Key Outcomes

  • Processed about 512 repositories into 37 structured service wikis.
  • Teams estimated engineer onboarding was about 40% faster and questions to senior engineers about 30% fewer.
  • Required every generated claim to cite a file path or be marked UNKNOWN.
  • Secured director sponsorship for a read-only employee knowledge-assistant proposal.

The Result

The repository agents produced 37 reusable service wikis from about 512 repositories. Team estimates put onboarding about 40% faster and support questions about 30% lower; the employee assistant remained a proposal, not a shipped outcome.