Groonga vs Sanity

Groonga ships small, precise releases roughly monthly, each documenting changes with runnable examples and crediting the reporter by name. The 16.0 line opened in February 2026 with an annual major release that deliberately carried no backward-incompatible changes. Recent work splits between conventional search functions — a new string_truncate(), vector support in between() — and a language model path where language_model_vectorize() consumes GGUF models from HuggingFace and TokenLanguageModelKNN handles query and passage prefixes. Contributor counts per release run to a handful of people.

The embedding functions are the notable thread. A search engine of this vintage adding model-backed vectorization as ordinary functions, callable from the same query language as everything else, is positioning for hybrid retrieval without a separate vector store in the stack. The rest of the cadence is characteristic of the project: careful compatibility work, Windows packaging problems chased across multiple releases, and build-system fixes contributed by downstream packagers.

Expect indexing support for the vector cases that currently work without an index — between() on vectors is explicitly flagged as unindexed — and further options on the language model functions. The project's no-breaking-changes stance on major releases suggests continuity rather than a rearchitecture.