An LLM that knows when to toast, when to flatter, and when to write your grandpa a fu

What it does

Tianji (天机, roughly “heaven’s secrets”) builds LLM applications around traditional Chinese social situations: toasting etiquette, gift-giving, defusing awkward moments, and composing blessings in styles ranging from casual chat to full classical Chinese prose. The same problem is solved four ways — prompt engineering, RAG knowledge bases, agents, and fine-tuned models — which is the point: it’s structured as a learn-by-example LLM development course.

The interesting bit

The “wishes” model (v0.2, in 3B/7B/14B Qwen-based sizes) writes personalized greetings in multiple registers — the sample output includes a Spring Festival message to grandpa written as a genuine parallel-prose fu. The real value, though, is the full pipeline being open: corpus, data-cleaning tools, training configs, and reproduction docs are all published on Hugging Face.

Key highlights

• Four implementation paths for the same task: prompt, RAG (LangChain/LlamaIndex), MetaGPT-based agents, and LoRA/full fine-tuning (Transformers, Xtuner)
• Open datasets on Hugging Face covering toasting, wishes, and general social-etiquette corpora
• Supports ChatGPT, Qwen, DeepSeek, ERNIE, InternLM, ZhipuAI — online or local
• Companion tutorial docs aim to take you from zero to entry-level LLM full-stack development
• Actively maintained through late 2024, with a WeChat community and multilingual READMEs (EN/JA)

Caveats

• The Agent module link in the README literally points to “修复中,请等待” (“under repair, please wait”) — that part is a work in progress
• Everything is Chinese-language and China-stack-centric (ModelScope, WeChat); less useful if that’s not your context

Verdict

A charming, unusually complete teaching repo if you read Chinese and want to learn prompt→RAG→agent→fine-tuning on one concrete problem. If you’re hunting for production etiquette tooling, this is a classroom, not a product.