Picking a vector store for fully offline RAG — filtering is the hard part
A discussion centers on choosing a for a system that must run completely offline, with no outbound at all. The core difficulty is at retrieval time — restricting results by document source, date, and permission level — which rules out bare-minimum . Qdrant handles this kind of filtering well, but its options for fully air-gapped use are awkward: the Hybrid Cloud version keeps an outbound connection open, the Private Cloud version requires a Kubernetes setup and a , and the open-source build lacks production-grade tooling.
LanceDB and Qdrant Edge are mentioned as newer alternatives worth considering. The post asks the community which people have settled on for isolated environments, how they handle without heavy , and whether going offline forced any compromise on filtering capability or retrieval accuracy (recall).
Key points
- (source, date, permission level) is the core challenge for offline RAG
- Qdrant filters well but its air-gapped deployment options are awkward
- Hybrid Cloud needs an outbound connection, Private Cloud needs Kubernetes plus a sales cycle, open-source build lacks production tooling
- LanceDB and Qdrant Edge are raised as alternatives worth checking out
- The post is an open question seeking community experience, not a benchmark or conclusion