The main topic was around how to choose a vector database, but we also discussed what is vector search, its use cases, aspects of building a multimodal search experience. I’ve also shown two demos, one built with FAISS (1M book titles), the other built with OpenSearch + GSI APU (10M images with captions), both sitting on top of Muves software stack.
Here is the recording:
We’ve got a bunch of questions from the audience — thanks to everyone for being so active and engaged during the event.
Amongst other key elements, common evaluation criteria include:
- Managed vs Self-Hosted: do you have own engineering team to look after self-hosting a vector database? If you do, you might consider hosting it with the help of your team, but it is also an opportunity cost by itself. If you don’t, having the vector database managed might be a good solution.
- Performance: what latency requirements does your app have? Is it served online or in batch offline mode?
- Existing MLOps: Machine Learning Operations (MLOps) includes the holistic view with tooling and processes for training and evaluating your ML models, getting feedback loop from production, accessing critical features for your model in cost-efficient way (like document features). If you have all of these in existence, then vector database might become one of the “modules” in your pipeline. If not (or not yet), then some of the operations could be handled by the vector DB, like computing embeddings for example.
- Developer experience: Availability of API docs, ability to customize components of the vector db (like a ranker), error handling, tech support — all these count, and you need to pay attention to dev experience, when choosing a vector db vendor.
- Reliability: what are the uptime SLA you find in the managed or self-hosted vector db? You don’t want to wake middle of the night when a node in your serving cluster goes down.
- Security: is this vector DB SOC2 compliant?
- Cost: how expensive is the cost of self-hosting / managed vector db. Again, include the opportunity cost, when your engineers are not developing the features you need in your product, while maintaining a database. On the other hand, having the opportunity to customize vector database to your needs might be critical for your business.
Hope you’ll enjoy this workshop recording, and remember to subscribe to Vector Podcast, where we discuss many of these aspects with the makers of vector search field: https://www.youtube.com/c/VectorPodcast