Data store bottleneck when grounding small document sets

[bendillinger]

Background
We want to use grounding our own data in a more on-the-fly way, similar to how NotebookLM notebooks can be created ad-hoc, but we’re running into 2 problems.

• Datastores take about 5 minutes to “spin up”, making it too slow for ad-hoc user interaction (and much slower than NotebookLM).
• There’s a limit of and 1M documents (which is plenty), but only 100 data stores. The 100 data store limit feels very constricting when each should hold about ~50 documents.

Questions

• If we want to use custom data grounding to build something similar to NotebookLM, are we thinking about datastores correctly?
• Should there be a 1-to-1 relationship between a datastore and notebook?
• Is there a way to subdivide the data in a datastore when using it for grounding so we can sidestep the spin-up time and datastore limit?