Hillock swaps the vector database for SQLite and SIMD hypervectors in local RAG

On October 3, 2026, a post on Reddit's LocalLLaMA community introduced a project called Hillock. The poster identified themselves as the project's creator, noting the disclosure up front and saying it was their first post in the community after spending time lurking and building up enough karma.

According to the author, every attempt to run local RAG on their own machine ran into the same two bottlenecks. The first was that standing up Chroma or another vector database alongside an 8B model, purely to chunk and parse documents, consumed VRAM that was needed for the main model. The second was that cosine similarity over text chunks performed poorly when the system had to refuse hard negatives.

Hillock tackles both by handling retrieval with SQLite plus SIMD supervectors rather than a dedicated vector database. The author says VRAM usage stays under 1.2GB.

The Reddit post is so far the only public account of the project. Beyond the memory figure, it offers no timeline, no third-party reproduction, no benchmark numbers, and no comparisons against other retrieval setups. Chroma and other vector databases appear only as the alternatives Hillock replaces. The author does not describe which platforms the project supports, how much data it is meant to handle, or under what license it is released.