This page describes strategies to implement multitenancy in Pinecone indexes. Multitenancy describes the act of keeping sets of vectors separate within a Pinecone index (e.g., segmenting vectors by customer, physically or logically). For more information about indexes, see Understanding indexes.

​
Namespaces

One approach to multitenancy is to use namespaces to isolate segments of vectors within a single index. This is a ‘pool’ model that shares most resources between tenants while keeping them logically separate.

​
Advantages

  • Namespaces reduce the need for additional indexes, reducing maintenance effort.

​
Disadvantages

  • Tenants share compute and storage resources.
  • Deleting tenant data is complicated and takes more time.

​
Metadata filtering

This approach to multitenancy stores all segments of vectors in a single index and filters by metadata at query time. This is another ‘pool’ model; here, you separate tenants at the query level.

​
Advantages

  • Metadata filtering allows you to query across multiple tenants.
  • Metadata filtering reduces the need for additional indexes, reducing maintenance effort.

​
Disadvantages

  • Tenants share compute and storage resources.
  • There is no way to track tenant-specific costs.
  • You cannot provision tenants with different dimensions, since dimensionality is set at the index level.

​
Separate indexes

Another approach to multitenancy is to create a separate index for each segment. This is a ‘silo’ model that provides dedicated resources to each tenant. For example, if you need to separate vectors for each customer, you can create a separate index for each customer.

​
Advantages

  • This model separates tenants physically, preventing queries across tenants and allocating compute and storage
    resources to each tenant.

​
Disadvantages

  • Creating and maintaining multiple indexes requires additional effort and cost.
  • This model doesn’t share resources between tenants, which is inefficient if you have tenants
    represented by a small number of vectors.
  • Creating a new index takes more time than creating a namespace.