The latest release of elstongun/leviathan is v0.3.0. This repository is a Rust-based single static binary that turns records exported from JSONL, JSON, CSV/TSV and SQLite, as well as from database CLIs, into a ranked full-text search index. Agents ask questions in natural language and receive relevant records as result cards with citations. What changed in v0.3.0 is not stated in the provided material.
The goal is to pull out the records you need like an agent's memory, instead of scanning huge local datasets every time. It uses SQLite FTS5 and BM25 for search. BM25 is a way of ranking documents by their relevance to the search terms. There is a CLI plus optional MCP and remote interfaces, and it also provides read/write memory features backed by a separate SQLite file. The license is Apache-2.0. The core of the design is gathering diverse records into a single index and returning them as citable results.
However, it is hard to tell from the README alone how the indexing approach differs from existing full-text search. The performance comparison is presented as a single measurement on one set of synthetic data the author created, so it is unconfirmed whether the same results hold on other data and workloads. With the current material, the search structure and features can be examined, but it is difficult to judge which one has the better search quality.