Collection titles are not unique, preventing tenant-based separation and lookup in VectorAPI

[jazzm0]

Describe the Problem/Bug

While working with the VectorApi and performing CRUD operations on collections, I noticed that collection titles are not required to be unique. This creates a significant challenge for our use case, where we intend to separate tenants by assigning each tenant its own collection.

Since titles are not unique, we currently have no reliable way to:

Query a collection by its title

Check whether a collection with a given title already exists

Current Workarounds

The only possible workarounds at the moment are:

Retrieve all collections and manually iterate over them

This becomes inefficient and error-prone as the number of collections grows.

Store a tenant → collectionId mapping on our side

This is not ideal, especially since the create-collection operation does not return the collection ID, which is already tracked in issue #672:
#672

Steps to Reproduce

Although this may not be a “bug” in the strict sense, it is a conceptual limitation that prevents us from using the VectorAPI for tenant isolation. Without unique titles or a way to query by title, it is very difficult to manage collections in a multi-tenant environment.

This is currently a blocker for our implementation.

Expected Behavior

We would greatly appreciate feedback on how collection identity is intended to work and whether one of the following enhancements could be considered:

Enforcing unique collection titles

Allowing lookup by title (e.g., GET /collections?title=)

Returning the collectionId when creating a collection

Adding support for queryable metadata to identify collections

Screenshots

No response

Used Versions

• Java and Maven version via mvn --version: ...

• SAP Cloud SDK for AI version: ...

• Spring Boot or CAP version: ...

  Dependency tree via mvn dependency:tree

Dependency tree here

Code Examples

// Your code here

Stack Trace

No response

Log File

Log file

...

Affected Development Phase

Getting Started

Impact

No Impact

Timeline

No response

[MatKuhr]

Tenant-specific content on AI Core is typically separated by resource groups. Is there a specific reason that doesn't work for your use case?

[jazzm0]

Hello,

this was, what Michael Haas recommended for us. Is it possible to create a resource group with an API? If yes, which one would that be? I searched through the sdk but didn't find the right one?

Kind Regards,
Tibor

[MatKuhr]

Here you go:

• https://github.com/SAP/ai-sdk-java/blob/main/core/src/main/java/com/sap/ai/sdk/core/client/ResourceGroupApi.java
• https://api.sap.com/api/AI_CORE_API/resource/Resource_Group

Please note, for grounding to work you must pass:

"labels": [
    {
      "key": "ext.ai.sap.com/document-grounding",
      "value": "true"
    }
  ]

as documented here

[jazzm0]

Hi Mathias,

this is probably a solution for the tenant separation problem, but this is just partially addressing the issue I mentioned. What about the query by title?

Kind Regards

[Jonas-Isr]

Hi @jazzm0,

if I understood you correctly, you wanted to have a way to distinguish collections with the same title from different tenants/users. This can be done with the information provided by Matthias above by assigning each tenant a specific resource group so that collections with different titles are in different resource groups and therefor easier to manage: You can then query in the relevant resource group to find the title you are looking for. If that is not a solution that works for you, can you elaborate more on your exact use case?

Best,
Jonas

[jazzm0]

Hi Jonas,

the idea is, that we do tenant separation based on resource groups. Since we will have different "sources" for the grounding we want to further distinguish which kind we want to use to give more specific answers. For this we planned to use collections per kind. That means, we need to be able to check wether a collection exists, to put the data there, or the query only for a particular collection. The problems with that I described above. Maybe we should have a call, as I said I'm also an SAP employee.

Kind Regards

[jazzm0]

Also, when reading the documentation I see the following he maximum number of resource groups is limited at tenant level to 50. If you reach this limit, you receive an error message. To free up space, delete some resource groups. Alternatively, raise a ticket to increase your quota. If we plan to do tenant separation (non BTP application) then 50 is way to low. We deal currently with thousands of tenants per customer. So creating a resource group for each tenant will not be feasible.

[MatKuhr]

The RG limit of 50 can be increased. Currently, this is a ticket-based process, we are working on improving that.

Please create a ticket for CA-ML-AIC and provide information about your product (your team info, link to cloud reporting showing the number of your tenants, link to your PPMS). Furthermore, please provide how many tenants you expect per BTP landscape and what the RG quota is you are going to need.

[Jonas-Isr]

Also, I want to add that this is not a bug or problem of the SDK but an issue with AI Core. So if there are no additional specific topics regarding the SDK in particular, I will close the issue and would ask you to proceed with the ticket as Matthias described above.

[MatKuhr]

With tenants being isolated via RGs, I'd expect the amount of collections you have to reduce significantly. So I'd think you can just get all collections and filter out the relevant one(s) from the resulting list. You can also assign arbitrary metadata to collections. And you can filter based on that metadata during retrieval calls. So I think with those tools I'd think you can organize the collections as you see fit.

If this is not the case, please outline what is missing.