# Caelune

A local-first Windows desktop application and read-only MCP server for searching private Markdown, PDF and Tika-backed document collections using a combination of exact text search and semantic vector search.

Caelune builds searchable indexes from documents already stored on the user's own computer, combining exact-text retrieval, structural signals, semantic vector search, optional reranking and source-aware context assembly. The desktop application manages data roots, runtime components, models, index builds, live file watching, queries and result review; a Query Console keeps input, scope and retrieval settings together, and a separate Results page shows source-labelled hits with snippet review.

An optional MCP server exposes the same retrieval core, through two read-only tools, to MCP-capable AI clients, but it can only search an index that the desktop app has already built, not create or modify one. Retrieval combines SQLite FTS5, structure-aware scoring, LanceDB vector search using the BAAI/bge-m3 embedding model, and an optional BAAI/bge-reranker-v2-m3 reranking step, and understands Markdown as well as common Logseq properties, block references and embeds.

All indexes, configuration, models, logs and runtime data stay under the active local data root. It targets Windows, is written in Python, and is MIT licensed.

Website: https://ellismorrow.github.io/Caelune/