Audit Finds 62% of 1,751 Relationship Milestones Extracted by RAG Are Junk

A team audited 1,751 relationship‑milestone entries generated by a Retrieval‑Augmented Generation (RAG) system. The audit revealed that 62 % of the extracted milestones were classified as junk. Junk entries included irrelevant text, duplicates, or nonsensical data. The findings highlight quality‑control challenges in automated content extraction. Researchers suggest improving filtering and validation steps to raise signal quality. The analysis was documented in a case study on the Astra platform. The report underscores the need for stricter evaluation of RAG outputs. Stakeholders are advised to monitor extraction pipelines for similar issues.