Why Tracking File Count Alone Causes Silent Rot in Semantic Search Indexes

A developer discovered that their semantic search index was returning confidently outdated answers because the re-embedding job only triggered when the file count changed, not when file contents were edited. This meant that any time a wiki article was updated without adding or deleting a file, the old vector embedding remained active in the database, matching queries as if the edit never happened — a failure the author terms a 'zombie vector.' The count-based trigger also never retired old chunks when files were split or merged, leaving 97 orphaned vectors pointing to source content that no longer existed. Over time, the index accumulated both stale embeddings and ghost references, causing an AI agent to retrieve outdated or broken information with no errors or warnings. The core lesson is that a reliable semantic index must track content changes — not just file count — to maintain the contract that vectors accurately reflect their source documents.