[All work](https://boiko.ai/work/)

Multimodal search · Desktop app

# Tern

Search a video archive by what was said, shown, or written on screen. Built to run locally on Apple Silicon.

[Explore the repository](https://github.com/B0yko/tern)

---

## [Find the moment, not the filename.](https://boiko.ai/work/tern/#find-the-moment-not-the-filename)

Tern brings speech, on-screen text, and visual search into one interface for audio, video, and photo archives. A query returns timestamped moments that can be previewed, trimmed, and exported.

## [Three ways into the same archive.](https://boiko.ai/work/tern/#three-ways-into-the-same-archive)

Whisper transcribes speech, Apple Vision extracts text, and SigLIP-2 embeds video keyframes. SQLite FTS5 and ChromaDB power the text and visual retrieval; the results are combined into a single ranked list.

Media and search indexes stay on the Mac. A FastAPI service connects the indexing and search pipeline to a vanilla JavaScript interface inside a Tauri desktop shell.

## [Making retrieval useful.](https://boiko.ai/work/tern/#making-retrieval-useful)

Matching results are grouped into time windows, with a ranking bonus when multiple search channels agree. The visual channel uses relevance thresholds to avoid returning confident-looking results for queries with no useful match.

A built-in trim editor turns search results into clips. Export options include MP4, MP3, subtitles, CSV, and FCPXML timelines for Final Cut Pro.

## [Where it stands.](https://boiko.ai/work/tern/#where-it-stands)

Tern is an unreleased v0.1 project. The public repository documents the implementation, tests, and remaining work. Its source is available for review under an all-rights-reserved license.

## [Related work](https://boiko.ai/work/tern/#related-work-heading)

- [wellbrief: offline report retrieval](https://boiko.ai/work/wellbrief/)

- [taskdistill: small local models](https://boiko.ai/work/taskdistill/)

Source: [https://boiko.ai/work/tern/](https://boiko.ai/work/tern/)