Vector Databases
Explore the fundamentals of vector databases used in Retrieval-Augmented Generation pipelines. Learn to select appropriate backends such as FAISS, Chroma, or managed vector databases based on persistence, metadata filtering, and scaling needs. Understand local versus managed storage trade-offs and the importance of schema design for filtering to build robust, scalable semantic search systems.
We need fast semantic lookup over a few thousand to a few million chunks of our internal docs, queries must support metadata filters, results must persist across restarts, and setup must stay simple on a laptop today while leaving a path to managed scaling later. The similarity math stays the same across all options, but the backend changes what gets indexed, what gets stored, which filters run at query time, and who owns the operational work when something fails.
The decision comes down to a few dimensions that we can check in code and in deployment.
A useful way to verify a choice is to pick one query we care about, run it end to end, restart the process, and run it again with the same filter. If results disappear, filters break, or index build time dominates startup, the backend is not matching the constraint set.
Local-first choices behave differently
With local-first storage, our process usually owns the ...