Vector Index Health Check

RAG quality degrades silently over time as duplicates accumulate (the same FAQ reindexed from three sources), as source docs get deleted but chunks don't, and as content becomes stale without being refreshed. Users see 'worse answers' and don't know why. A periodic audit surfaces the cause instead of waiting for complaints.

1. Pull metadata for every vector — source URL, last_modified, chunk_id. 2. Near-dup scan: for each chunk, get its top-5 nearest neighbors; flag any pair with similarity > threshold where the text diff is only whitespace + boilerplate. 3. Orphan scan: resolve each source URL (HEAD request) and flag 404s or redirects to different content. 4. Stale scan: flag chunks whose source last_modified exceeds the threshold. 5. Emit a markdown report with counts, example entries, and suggested remediations (dedup, delete, refresh).