Enhanced Agentic Rag

Uber Engineering reports improving the answers of Genie, its internal on-call assistant, for engineering security and privacy questions by repairing both document processing and retrieval. Its engineers preserved table structure when reading policy documents, enriched them with summaries, FAQs and keywords, and added agents that clarify questions, select likely policy documents and organize retrieved passages before an answer is written. These steps address a concrete failure: policy passages are often similar but differ in details such as retention rules across personas and geographies, so simple similarity search can return the wrong one.

On a test set of more than 100 security and privacy queries curated by subject-matter experts, Uber reports a 27% relative increase in acceptable answers and a 60% relative reduction in incorrect advice, compared with its earlier traditional retrieval-augmented setup. Those results describe the combined system changes on that internal workload. Architecture and evaluation account