I find the Pinecone assistant super useful to start out with a use case. Great job!
What I dearly miss is the possibility to test namespaces with the assistant. Example: If my use case involves three clients I would want to be able to separate client specific documents that I upload. Namespaces would be the obvious choice. Please provide in the upload workflow a way to indicate the namespaces.
Secondly, if I am happy with the results from the assistant I would like to scale it with the vector database itself. However, all the heavy lifting that the assistant took away with its optimized automatic handling of documents, i.e. chunking, embedding with a suitable model, vector index management, etc. is not replicable since the configuration parameters are not accessible. Please provide an overview of these parameters so it can be replicated at scale and yield the same results.
That would make the assistant even more productive and I am sure would lower the barrier for more people with more use cases.
1 - Regarding namespaces, it may actually make more sense for you to either a) create separate assistants for your different clients or b) use metadata filtering on the files to upload everything into a single Assistant and filter at query time
2 - The Pinecone Assistant itself is intended to operate at scale and continue taking away all that “heavy lifting” for you. Simply sharing those config parameters doesn’t not equate to you being able to actually replicate the quality and functionality that the Assistant is bringing. Do you have any particular concerns over its scalability?
1 - Pinecone advocates namespaces as the best solution to segregate clients and/or documents with restricted access. My current use case is to separate country documentations that deviate only slightly and each country team should only have access to the respective documentation. So, I would like to use namespaces to be super sure. Multiple assistants are not an option given the many countries we serve, since I want to minimize maintenance.
2 - It appears to me that the assistant is more intended for a quick deployment and testing of an use case. Happy to hear that is is designed to operate at scale. However, I see a lot less endpoints that I can use with the assistant. I can e.g. not upsert a record or enforce security options. Therefore I would like the good parts of both products (assistant’s ease of use in generating and querying the vectors and database’s professional features) to be available in one product.
Thanks again and really appreciate your continued feedback.
1 - Yes, understood that on the vector database we recommend namespaces…as Assistant is a higher abstraction we do recommend using multiple Assistants. We have production customers with 10s of thousands of Assistants running today. Since it is all API driven, there should really be no difference in how you manage multiple namespaces vs multiple Assistants.
2 - We hear you loud and clear on this. As I mentioned, the Assistant is meant to be an abstraction layer above the DB, intended to operate on whole “documents” rather than individual vectors/chunks. However, we are also exploring how to get the best of both worlds and your feedback is very much inline with that
I’m curious what security options you’re missing with Assistant?
