Run qmd cleanup. This performs 4 steps sequentially:
- Delete all LLM cache entries from the llm_cache table (cached query expansion and reranking results from previous searches) — reports count cleared
- Remove orphaned vectors (embeddings with no corresponding document reference — vectors for deleted or changed documents) — reports count removed or "No orphaned embeddings to remove"
- Delete inactive documents (records with active=0, soft-deleted during indexing when files are removed or modified) — reports count removed if > 0
- Run SQLite VACUUM to reclaim disk space from deleted records — reports "Database vacuumed"
Each step shows a green checkmark on completion. Note that searches will be slower after cleanup until the LLM cache rebuilds through natural query usage.