Can you share a specific example of a complex failure you encountered with this setup and how you resolved it?

In a production RAG pipeline, we observed that queries about financial metrics returned outdated tables because the vector store had stale embeddings. The failure manifested as low precision and user complaints. I first reproduced the issue locally by querying the same vector store and inspecting the top‑k results. I discovered that the embeddings were generated with an older model version. To fix it, I re‑indexed the entire corpus with the updated embedding model and added a metadata field for the document’s last‑updated timestamp. I then implemented a hybrid search that first filters by timestamp (e.g., only documents updated in the last 30 days) before performing vector similarity. After redeploying, I monitored precision@1 and saw a 15% improvement. I also set up a scheduled re‑index job and added a monitoring alert for embedding drift to prevent future regressions.