SlothChat：可離線使用、採用開放權重模型的羣組聊天室摘要工具

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady.

What I Built

SlothChat is a local-first, privacy-focused AI tool that reads your chaotic group chats overnight and hands you a clean, 3-bullet summary with action items in the morning.

We all have that one group chat—the one with 150 unread messages about inside jokes, server crashes, and "who's bringing what." Instead of scrolling for 20 minutes to catch up, SlothChat does the heavy lifting. You simply export your WhatsApp chat as a .txt file, drop it into the app, and the AI extracts the signal from the noise.

Code

GitHub Repository: https://github.com/Dexter-6666/slothchat

How I Built It

The core of SlothChat is a local inference pipeline running the Qwen 2.5 1.5B model via HuggingFace transformers.

Instead of sending your private group chats to a closed API, the entire process runs on your machine (or a free Colab instance). The script uses a custom regex parser to strip WhatsApp timestamps and format the text into a strict prompt. I pinned the temperature to 0.1 to force the model to stick strictly to the facts, avoiding creative hallucinations.

Handling the Context Window

A deliberate technical choice I made was processing the last 100 messages. While open-weight models can handle massive context windows, feeding a 5,000-message chat into a 1.5B parameter model results in the "Lost in the Middle" phenomenon (where the model forgets details buried in long text) and massive latency. By focusing on the most recent 100 messages (roughly 1,000 tokens), SlothChat generates an accurate, lightning-fast summary of exactly what you missed overnight.

Why Does Open Innovation Matter?

Overnight group chats are the absolute worst thing to paste into a closed, rented model. It contains sensitive context, personal names, and half-baked plans.

Running an open-weight model locally changes that entirely:

1. Privacy: The chat stays on your laptop. It never becomes someone else's training data.
2. No Metering: You don't pay per token. You can summarize 50 chats a day and it costs you exactly $0.
3. Offline Capability: It works on a plane, in a cabin, or during an internet outage.

The Lazy Payoff

SlothChat embodies the sloth lifestyle. It sits quietly, consumes zero cloud resources, and does the heavy cognitive lifting so I don't have to. I've traded the anxiety of a red notification badge for a clean, 3-line summary.

Tech Stack: Python, HuggingFace Transformers, Qwen 2.5 (Open-Weight), PyTorch, Gradio.