In addition to upserting and querying data, there are other ways you can interact with vector data in a Pinecone index. This section walks through the various vector operations available.

​
Create a client instance

If you’re using a Pinecone client library to access an index, you’ll need to create a client instance:

import pinecone

pinecone.init(api_key="YOUR_API_KEY", environment="YOUR_ENVIRONMENT")
index = pinecone.Index("pinecone-index")

​
Specify an index endpoint

Pinecone indexes each have their own DNS endpoint. For cURL and other direct
API calls to a Pinecone index, you’ll need to know the dedicated endpoint for
your index.

Index endpoints take the following form:

https://$INDEX_NAME-$PINECONE_PROJECT_ID.svc.$PINECONE_ENVIRONMENT.pinecone.io

  • $INDEX_NAME is the name you gave your index when you created it.
  • $PINECONE_PROJECT_ID is the Pinecone project id that your API key is associated
    with. This can be retrieved using the whoami operation below.
  • $PINECONE_ENVIRONMENT is the cloud region for your Pinecone project..

​
Call whoami to retrieve your project id.

The following command retrieves your Pinecone project id.

pinecone.whoami()

​
Describe index statistics

Get statistics about an index, such as record count per namespace:

index.describe_index_stats()

​
Fetching records

The Fetch operation looks up and returns records, by id, from an index. The returned records include the vector values and/or metadata.

Fetch records by their ids:

index.fetch(["id-1", "id-2"])

# Returns:
# {'namespace': '',
#  'vectors': {'id-1': {'id': 'id-1',
#                       'values': [0.568879, 0.632687092, 0.856837332, ...]},
#              'id-2': {'id': 'id-2',
#                       'values': [0.00891787093, 0.581895, 0.315718859, ...]}}}

​
Updating records

There are two methods for updating records and metadata, using full or partial updates.

​
Full update

Full updates modify the entire record, including vector values and metadata. Updating a record by id is done the same way as inserting records. (Write operations in Pinecone are idempotent.)

The Upsert operation writes records into an index.

If a new value is upserted for an existing vector id, it will overwrite the previous value.

  1. Update the value of the record ("id-3", [3.3, 3.3]):
index.upsert([("id-3", [3.3, 3.3])])
  1. Fetch the record again. We should get ("id-3", [3.3, 3.3]):
index.fetch(["id-3"])

​
Partial update

The Update operation performs partial updates that allow changes to part of a record. Given an ID, we can update the vector value with the values argument or update metadata with the set_metadata argument.

The Update operation does not validate the existence of ids within an
index. If a non-existent id is given then no changes are made and a 200 OK
will be returned.

To update the vector values of record ("id-3", [3.0, 3.0], {"type": "doc", "genre": "drama"}):

index.update(id="id-3", values=[4.0, 2.0])

The updated record would now be ("id-3", [4.0, 2.0], {"type": "doc", "genre": "drama"}). Values have been updated but the metadata is unchanged.

When updating metadata only specified fields will be modified. If a specified field does not exist, it is added.

Metadata updates apply only to fields passed to the set_metadata
argument. Any other fields will remain unchanged.

To update the metadata of record ("id-3", [4.0, 2.0], {"type": "doc", "genre": "drama"}), use code like the following:

index.update(id="id-3", set_metadata={"type": "web", "new": True})

The updated record would now be ("id-3", [4.0, 2.0], {"type": "web", "genre": "drama", "new": true}). The type metadata field has been updated to web, the new property has been added with value true, and the genre property has been unchanged.

Both vector and metadata can be updated at once by including both values and metadata arguments. To update both these parts of the "id-3" record we write:

index.update(id="id-3", values=[1.0, 2.0], set_metadata={"type": "webdoc"})

The updated record would now be ("id-3", [1.0, 2.0], {"type": "webdoc", "genre": "drama", "new": true}).

​
Deleting records

The Delete operation deletes records from an index.

Alternatively, it can also delete all records from an index or namespace.

When deleting large numbers of records, limit the scope of delete operations to hundreds of records per operation.

Instead of deleting all records in an index, delete the index and recreate it.

​
Delete records by ID

Example

index.delete(ids=["id-1", "id-2"], namespace='example-namespace')

​
Delete records by namespace

To delete all records from a namespace, specify the appropriate parameter for your client and provide a
namespace parameter.

If you delete all records from a single namespace, it will also delete the namespace.

Example:

index.delete(delete_all=True, namespace='example-namespace')

​
Delete records by metadata

To delete records by metadata, pass a metadata filter expression to the delete operation.