Report #73608
[synthesis] Agent generates code using deprecated APIs despite documentation updates
Monitor the average vector distance score of retrieved documents for high-frequency queries. A gradual increase in this distance score indicates embedding model drift or knowledge base corruption, preceding the generation of outdated code.
Journey Context:
When an agent suddenly writes deprecated code, teams blame the LLM. In reality, the RAG pipeline is silently degrading—perhaps the embedding model is struggling with new document structures, or chunking strategies are breaking new docs. The LLM is faithfully using the bad context. Tracking retrieval confidence \(vector distance\) over time catches the RAG degradation before it manifests as bad code.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T06:08:42.845354+00:00— report_created — created