My Friend Can't Explain Loglan Grammar Fast Enough — So I Built Him an AI That Speaks a 1955 Artificial Language

Описание и применение

Алекс, наставник по логлану, тратил 20–30 минут на ответы на грамматические вопросы, используя большие базы данных. Созданный локальный LLM‑ассистент, основанный на двухуровневом RAG‑системе, отвечает за миллисекунды, предоставляет точные ответы с цитатами и устраняет галлюцинации.

Задача

Alex, a Loglan mentor, spends 20–30 minutes answering grammar questions that require searching a 10,000‑word SQLite database and cross‑referencing textbooks, and generic LLMs hallucinate words, borrow from Lojban, and mis‑assign argument slots.

Результат

The assistant reduces answer time from 20–30 minutes to milliseconds, provides accurate, citation‑backed explanations, and achieves near‑zero hallucination rates by grounding all context in the LOD dictionary.

Рабочий процесс

• 1. Ingest 28 canonical Loglan sources into an SQLite FTS5 index (18,766 dictionary definitions and 769 textbook chunks).
• 2. When a user queries, the retriever searches Tier 1 (LOD dictionary) and Tier 2 (grammar corpus) simultaneously.
• 3. The grounded context is fed to a local Ollama model.
• 4. The model outputs a markdown answer with slot definitions, parse tree, and citations.
• 5. The assistant is available via a terminal REPL and a Streamlit web demo.

Доказательства

Each answer requires digging through a 10,000-word SQLite database, cross-referencing 1970s grammar textbooks, checking case-tag theory papers, and formulating a coherent explanation. It is mentally exhausting and takes 20–30 minutes per question.

Источник

A standard cloud model fails badly on Loglan: It hallucinates words that don't exist in the 9,988-word dictionary, borrows from Lojban, and scrambles argument slots.

Источник

Grounding with the LOD dictionary fixes this: when the retriever explicitly surfaces cinkau → puppy / infant-dog, hallucination drops to near zero.

Источник

I used to spend 15 minutes explaining how 'nu donsu' flips the donor and the gift, and another 10 minutes digging up the lesson on case tags. Now I just paste the assistant's breakdown into our community channel.

Источник