zvec

Nice project. The 4GB index / ~15s search part made me think zvec might be a good fit here: https://github.com/alibaba/zvec It’s an in-process vector DB, so the “local corpus, no separate server” shape is pretty much what it’s designed for. Its benchmark numbers are quite strong, and recent versions also support full-text + hybrid retrieval and DiskANN. This would be an interesting case to try with zvec: same corpus, same embedding model, then compare indexing time, index size, memory usage, and query latency on normal hardware.

There's also the recently released zvec[1], the tagline for which is: The SQLite of Vector Databases. [1]: https://github.com/alibaba/zvec

Zvec is a lightweight, SQLite-like embedded vector database purpose-built for edge and on-device workloads. It delivers four core advantages: out-of-the-box usability, configurable resource budgets, extreme performance, and versatile vector capabilities. Open-sourced under the Apache 2.0 license, Zvec aims to provide developers with a one-stop path from prototyping to production deployment on the edge. Try Zvec—we’re excited to see your adoption, feedback, and contributions. For a more detailed introduction, you can refer to https://zvec.org/en/blog/introduction/