Zakonnik BY

People ask legal questions in everyday language, but the source material is written in legal language and scattered across long codes. A general-purpose chatbot answers confidently and gives no way to check whether the answer came from an actual article of law.

Built for non-lawyers who need a first orientation in Belarusian law — what article applies, roughly what it says, and where to read it in full. It is an orientation tool, not legal advice.

• →Split the legal corpus into retrievable chunks that stay aligned with article boundaries, so a citation points to something a person can actually open.
• →Put retrieval in front of the model instead of relying on the model's own memory of the law.
• →Constrain the prompt layer so an answer without a supporting source is refused rather than improvised.
• →Ship it inside Telegram, because that is where the audience already is — no new app to install or account to create.

1. 01User question in Telegram
2. 02Retrieval over the chunked legal corpus
3. 03Prompt layer: question + retrieved articles + refusal rules
4. 04Model answer with citation back to the source article

• Telegram Bot API
• StepFun
• Retrieval / RAG pipeline
• Prompt + citation layer

• +Article-aligned chunking: citations stay verifiable because a chunk maps to a real article.
• +Answering in plain language while keeping the legal reference visible next to it.
• +Telegram as the runtime — zero onboarding friction.

• −Naive fixed-size chunking cut across articles and produced citations that pointed at nothing useful.
• −Without explicit refusal rules the model filled gaps with plausible-sounding law.

Retrieval, prompt layer and bot runtime were designed, tested and shipped end to end. It runs as a working Telegram assistant with citation-backed answers.

In a legal RAG system the retrieval boundary is the product. How you cut the corpus decides whether a citation is trustworthy — the model choice matters far less.