Implementing RAG on AWS: do you store document chunks alongside your vectors or in separate systems (e.g., S3)? What methods do you use for storage and metadata linkage?
yeye, in my own experiance i lean towards storing docs separately. vec db rocks for quick lookups, while s3 or similar is perfect for large doc chunks. theres a bit extra linking work but its worth it for scalability
im all for storing docs separately. vecdb is super fast for queries and s3 handles heavy docs better. sometimes linking metadata feels messy but its less of a headache than trying to mix both storage types.
In my experience implementing RAG on AWS, the choice depends significantly on priorities such as scalability, operational complexity, and cost management. I have found that maintaining flexible metadata linkage is easier when document chunks are stored separately in a system like S3, especially when dealing with large documents and frequent updates. Consolidating data can lead to performance benefits for low-latency applications, but separating storage allows for better resource allocation and independent scaling. Ultimately, ensuring robust identifier mapping is key in either approach to maintain data consistency and retrieval efficiency.
hey, i’ve been mulling over doc storage too. mixing vecdb retrieval with separate docs seems safer to me. curious if anyone’s tried a hybrid setup? how do you think metadata updates fare with that approach?
In my view, leveraging a dual-system approach where vector searches and document storage are handled separately offers an effective balance between speed and maintainability. By utilizing vector databases for rapid similarity searches and systems like S3 for comprehensive document storage, one achieves scalability without sacrificing metadata integrity. This method simplifies handling updates and ensures data consistency across systems. My practical experience confirms that separating the two components reduces operational complexity, particularly during metadata synchronization, thereby optimizing system performance and maintenance in production environments.