Back up, move and restore a collection¶
A local collection is a handful of files under its path: the vector index, the SQLite stores for texts and metadata, and, in graph mode, the knowledge graph. export() packs all of them into one zip with a manifest; import_snapshot() unpacks it and opens it with the same mode and models.
Export¶
from vectrixdb import Vectrix
db = Vectrix("notes", path="./data")
db.add(["first note", "second note"])
snapshot = db.export("backups/notes.zip")
The collection is flushed first, so the zip holds committed state. The manifest records the name, mode, dense model, dimension, document count and the VectrixDB version.
Import¶
restored = Vectrix.import_snapshot("backups/notes.zip", path="./restored")
restored = Vectrix.import_snapshot("backups/notes.zip", path="./restored", name="notes-copy")
Mode and dense model come from the manifest; anything else you pass goes to Vectrix on top. Import refuses to overwrite a collection that already exists at the target, so a mistaken restore cannot clobber live data.
The format is deliberately plain: unzip the file and you have exactly what VectrixDB wrote, in the layout it expects.
Rebuild the index¶
Deletions leave tombstones in an HNSW graph, and recall drifts as the graph is edited in place. A rebuild starts clean from the live vectors:
The new index is built outside the lock, so searches keep running against the old one until the swap, which is atomic.
Deletions cost speed as well as recall, because a deleted vector stays in the graph and is filtered out after the search rather than skipped during it. Measured on 20,000 vectors: with a fifth of them deleted a query went from 1.0 ms to 18.0 ms. You do not have to watch for it. Collection.tombstone_ratio reports the share, and past a fifth search() says so on the result:
results = db.search("volcanic rock")
if results.degraded:
print(results.degraded)
# 30% of this index is deleted vectors, which slows every search;
# call rebuild_index() to compact it
Open read-only, larger than memory¶
The index is memory-mapped instead of loaded, so opening is instant and a collection larger than RAM can still be searched. Any write raises ConfigurationError. Combine with export() to ship a built collection to machines that only query it.