I built a notes search that never phones home (23 MB model, one JSON file)

Description and use

slothbox is a local semantic search tool for notes that uses an open embedding model, stores its index in a single JSON file, and operates entirely offline.

Problem

Existing note‑search products require sending user text to external servers, an API key, and a credit card, making it costly and privacy‑concerned for personal note collections.

Outcome

slothbox builds a 23 MB model index in ~3 s, embeds 5 files in under 2 s, produces ~60 KB JSON, and makes zero network calls after the first run, enabling fast, private, unmetered note search.

Workflow

• 1. Run node sloth.mjs index ./notes to build the index.
• 2. Query with node sloth.mjs ask "question" to retrieve relevant passages.
• 3. The tool skips unchanged files, saves progress after each file, and sleeps 80 ms between files to run in background.

Evidence

Every "ask your notes" product I tried wanted the same three things: my text on their server, an API key with a monthly ceiling, and a credit card.

Source

The whole index is one JSON file next to your notes.

Source

After that there are zero network calls. I verified this the boring way: ran the same query with the Hugging Face hub disabled (HF_HUB_OFFLINE=1) and got the same answer, with no network access.

Source

Model load, first call ~3.1 s; Embed 5 files / 5 passages under 2 s; Store size, 5 passages ~60 KB of JSON; Network calls after the first run 0.

Source

Retrieval is not understanding. Passage-level cosine similarity finds the right paragraph and then stops.

Source