Nhakogithub.com/kimzam30/Nhako-BotSelf-host
Nhako Bot
A Telegram assistant that runs entirely on your own hardware: a local LLM, local vector memory and offline voice transcription.
- Status
- Self-host
- Kind
- Telegram assistant
- Model
- Llama 3.2 via Ollama; nomic-embed-text for embeddings
- Memory
- PostgreSQL with pgvector, retrieved by cosine distance
- Speech
- faster-whisper on CPU (int8)
- Runs as
- Python 3.11+, or a Docker container
- Started
- Last push
- License
- MIT
- Built by
kimzam
What it does
| Capability | How |
|---|---|
| Chat | Llama 3.2 through Ollama: private, no API bill, no rate limit |
| Long-term memory | Conversations embedded and stored in pgvector; relevant history is retrieved and added to the prompt |
| Voice notes | Transcribed offline with faster-whisper |
| Web search | DuckDuckGo on demand, so answers are not capped at the model's training cutoff |
| Calendar | Reads, creates and deletes Google Calendar events from plain language |
How it works
A small router inspects each incoming message and sends it to the right tool before handing context to the model, rather than trusting the model to choose. Nothing is sent to a hosted AI provider.
What is next
- Scraping links with automatic tagging
- Vision through
llama3.2-vision - Homelab monitoring commands: container status, disk and uptime