Working with collections¶

A collection is a read-only, static snapshot of a pod-based index’s vector data, held outside the index. Collections and backups are two different mechanisms for two different deployment types, and they are not interchangeable:

	Snapshot mechanism	Taken with	Restored with
Pod-based index	Collection	`pc.collections.create()`	nothing — see below
Managed (serverless) or BYOC index	Backup	`pc.backups.create()`	`pc.create_index_from_backup()`

For the backup side, see backups and restore.

Important

API version 2026-07, the version this SDK targets, does not create pod-based indexes, so there is no source pc.collections.create() can point at. Everything else on this page still works: the /collections routes are served, pc.collections and AsyncPinecone.collections are still on the client, and list(), describe(), and delete() operate normally on collections that already exist. See pod-based indexes for the refusal itself, and Pod deployments, and what that means for collections in the v10 migration guide for the API-version pin that reaches an earlier version.

Create a collection¶

create() takes a name for the snapshot and the name of the pod index to copy:

col = pc.collections.create(name="movie-embeddings-snapshot", source="movie-recommendations")
print(col.status)   # "Initializing"

The source has to be a pod index that is already ready; a managed source is rejected. The call returns as soon as creation starts, and there is no timeout= argument to wait on — poll describe until the status leaves "Initializing".

Because 2026-07 creates no pod index, this call has no reachable source there. For snapshot and restore on managed indexes, use backups instead — see backups and restore, and pc.create_index_from_backup(...) for the restore path.

List collections¶

list returns a CollectionList you can iterate or call .names() on. There is no filtering, sorting, or pagination — every collection in the project comes back at once:

for col in pc.collections.list():
    print(col.name, col.status)

names = pc.collections.list().names()
print(names)   # ['movie-embeddings-snapshot', 'product-catalog-snapshot']

This is the way to inventory the collections a project already holds, including before an admin.projects.delete(), which a leftover collection blocks with FailedPreconditionError naming what is still there.

Describe a collection¶

describe returns a CollectionModel:

col = pc.collections.describe("movie-embeddings-snapshot")
print(col.name)          # 'movie-embeddings-snapshot'
print(col.status)        # 'Ready'
print(col.dimension)     # 1024
print(col.vector_count)  # 99
print(col.size)          # 3126700
print(col.environment)   # 'us-east1-gcp'

size is how much space the snapshot occupies in bytes, not a vector count — that is vector_count. size, dimension, and vector_count are all None until the collection finishes initializing, so read status before trusting them.

Delete a collection¶

pc.collections.delete("movie-embeddings-snapshot")

Deletion is asynchronous: the call returns as soon as the request is accepted, and the collection can still appear in list for a short time afterwards. The source index cannot be deleted until the collection is really gone.

delete raises NotFoundError if the collection does not exist.

Create an index from a collection¶

There is no path from a collection back to an index. 2026-07 rejects index creation from a collection, and the SDK refuses both spellings client-side rather than sending a call it knows will fail:

pc.indexes.create(source_collection=...) raises PineconeTypeError.
source_collection set on a PodSpec passed to the deprecated spec= argument raises the same error — the field has no destination in the current create request, and dropping it silently would send a different call than the one you wrote.

Both messages name pc.create_index_from_backup(backup_id=..., name=...) as the supported restore path. See backups and restore; it covers managed (serverless) and BYOC indexes, since pod indexes cannot be backed up either.