Stale embeddings silently degrade RAG quality: users see outdated policies, wrong pricing, or revoked guidance. Treat the embedding index as a production data product with SLAs, not a one-off batch job.

Ingestion design

• Use change data capture or webhooks from source systems instead of nightly full rescans when possible.
• Version chunks with content hashes so you only re-embed when text truly changes.
• Propagate legal holds and deletions into the vector store promptly to meet retention obligations.

Chunking and overlap

Balance recall and precision: tiny chunks miss context; huge chunks dilute relevance. Measure on real user questions, not toy benchmarks. Consider structure-aware splitting for HTML, Markdown, and PDF headings.

Quality loops

• Offline nDCG or MRR-style metrics on labeled question sets after each index rebuild.
• Online logging of which chunks appear in top-k for high-traffic queries.
• Human review queue for low-confidence answers tied to source document diffs.

When pipelines are observable and incremental, on-call can explain why an answer changed—and fix it without a panic full reindex.