I'm building an app similar to Duolingo and I'm considering using a vector database because I plan to integrate LLM features in the future. Right now, I'm exploring pgvector since I'm new to PostgreSQL, having primarily worked with MySQL. I've heard that pgvector could have memory issues and may take longer to process when working with large datasets. Is it a bad idea to start with a vector database or would it be wiser to stick with regular PostgreSQL at the beginning and add vector search later? Any advice from those with real experience would be greatly appreciated!
5 Answers
Vector and traditional databases serve different purposes; they're not the same. Unless you have a specific use case for vector searches right now, it's better to go with a standard database.
Financial considerations matter too—if you don't have a dedicated server for this, I'd suggest using two different databases for your features and starting with something like MariaDB or PostgreSQL.
Since you're seeking advice here, take things incrementally. Focus on creating a proof of concept first. You can always switch to another database or migrate your data as your project grows.
It's hard to get ahead of yourself with database planning. Without a defined feature or an architecture in mind that necessitates a vector database, it's safer to stick with standard SQL like PostgreSQL.
In general, databases designed for specific tasks often don't perform as well as established general-purpose options like PostgreSQL. It's usually best to start with a reliable, all-purpose database until you hit a specific scalability hurdle that justifies switching to something specialized. Planning ahead can lead to unexpected challenges, and most of the time, the issues you foresee aren't the ones that actually arise.
While that's mostly true, there are exceptions. For instance, time series databases have been shown to exceed PostgreSQL's performance for financial data. Also, columnar databases can outperform general-purpose dbs in BI scenarios due to their structure.