A vector index tombstone is a logical deletion marker that hides a removed record before the underlying search structure has physically reclaimed it.
In a local RAG index, deleting a PDF or photo can make its chunks disappear from normal results long before the vector graph, segment files, or storage pages are rewritten. Tombstones let the database preserve index consistency while maintenance happens later. That intermediate state matters for storage estimates, index health, update churn, and any assumption that a logical delete is the same as immediate physical erasure.
A Tombstone Marks a Record Deleted Before Physical Cleanup
Deleting a record from an approximate index cannot always remove every physical reference in one cheap operation. The system can instead mark the item as deleted, exclude it from visible results, and leave structural cleanup to a later maintenance pass.
In an HNSW index, tombstones mark deleted objects and can remain in the graph until a cleanup process removes them.
For a home knowledge base, this means the source file can be gone and the query layer can correctly suppress its chunks even though internal index state still remembers that those nodes once existed.
Search Can Hide a Deleted Vector While the Index Still Carries State
Logical deletion and physical reclamation answer different questions. The first asks whether the record is eligible to appear in search; the second asks whether its bytes and index relationships have been removed from storage and memory structures.
Segment-based indexes can leave deleted records until merge instead of forcing every delete to rewrite a segment immediately.
This separation is why a collection count can fall before disk usage does. It also explains why heavy delete/update workloads can accumulate maintenance debt without returning stale records to ordinary search.
A monitoring dashboard should therefore distinguish live records, pending deleted state, segment count, and actual disk usage instead of treating one metric as proof that cleanup has finished.
Tombstones Accumulate When Files Change or Are Replaced Repeatedly
A private index can generate deletion markers during normal updates, not only when a user permanently removes a file. Replacing one document version may delete old chunks and insert new ones, while reorganizing folders can retire records under previous identities.
Mutable vector collections accumulate updates and deletes before segment optimization consolidates those changes asynchronously.
If a household knowledge base re-embeds documents frequently, tombstone pressure can become a better indicator of churn than the number of currently searchable files. Cleanup frequency should reflect update rate and available I/O headroom.
A Tombstone Is Not Secure Erasure
A logical marker is designed for index correctness, not forensic deletion. Old bytes may remain in segment files, snapshots, replicas, backups, filesystem free space, or other derived stores until separate lifecycle processes remove them.
The later compaction after deletion stage is a different mechanism from the tombstone itself because it can rewrite obsolete segment state and reclaim space.
Treat sensitive deletion as an end-to-end retention problem covering source files, vector stores, metadata, caches, snapshots, and backups. Tombstones are one intermediate consistency mechanism inside that larger lifecycle.
This boundary also prevents a misleading storage expectation: deleting thousands of chunks can be immediately correct for search while remaining invisible in free-space graphs until scheduled cleanup finishes.
Tech & AI HUB
More to Read

What Is Plex State, and Which Parts Must Persist?
Persistent Plex state is the information that preserves the server experience across restart and rebuild; media and temporary transcode data are separate roles.

How Does Plex Handle Authentication Across Local and Remote Sessions?
Plex authentication starts with server and account identity, then local or remote network paths determine reachability and secure connection behavior.

Why Can Plex Search Slow Down as Library Data Grows?
Library growth alone is not the diagnosis. Test query shape, indexes, cache state, storage latency, and write activity before blaming database size.

