JEV Telegram Filter

A local Telegram conditional filter: FastAPI and SQLite listen on 127.0.0.1:18721, one Telegram account's messages are stored on your machine, and tasks send message batches to TypeSafe Jev or local Laya so that only hits are shown.

Each task batches messages (default batch size 50) and sends them to the analysis backend you pick — TypeSafe's hosted Jev, or local Laya. You set a hit threshold and tags per task, and the hits can be filtered by a minimum noul value.

Python 3.11+, uv, and Node.js 20+ for development; a Telegram api_id and api_hash from my.telegram.org; and for the Jev backend a TypeSafe API key entered in settings, with TYPESAFE_API_KEY as the environment fallback before a key is saved.

Yes: choose Laya as the analysis backend and it uses the local Ollama model laya (start Ollama and pull the model first) or Unsloth Desktop at 127.0.0.1:8888. The README says the Python laya package is not required, and weights are managed by that app, not this repo.

The analysis toggle pauses or starts analysis while messages keep arriving. Messages must fill a task's batch size and fall inside the days window to queue; hits appear in the right column with images above text, with server-side search and filters by task, type, and minimum noul. Misses are never listed.

Messages and settings stay in local SQLite under data/; packaged builds keep the same files under the per-user APPDATA directory on Windows and ~/Library/Application Support/jev_tg on macOS, overridable with JEV_TG_DATA_DIR. The README warns data/, .env, and session files are gitignored and must not be committed.

GitHub metadata lists no license for the repository. The release workflow builds jev-tg.exe for Windows and JEV-Telegram-Filter.dmg for macOS on pushes to main or master, tags vYYYYMMDD.HHMMSS, and publishes a GitHub Release; commits containing [skip release] skip that release.