Al-Muhami Al-Zaki
A Corrective RAG for Egyptian legal research that admits uncertainty instead of inventing an article number. Arabic semantic search with mandatory Law, Article and Year citations, running on local models.
Egyptian legal texts indexed for Arabic semantic search.
Mandatory on every answer.
Documents and questions stay on the machine.
Problem
Legal research is the worst place for a confident hallucination, and Arabic legal text is poorly served by mainstream tools. A wrong article number stated fluently is more dangerous than "I don't know."
Approach
A Corrective RAG pipeline on LangGraph. Retrieved passages are relevance-graded and validated before generation, and the query is rewritten when retrieval fails. Answers carry mandatory Law, Article and Year citations. CAMeLBERT-NER anonymises personal data before it's processed, and everything runs on local Ollama models.
Outcome
An early, working instance of the two things I now build professionally: cited, refuse-when-unsure RAG, and Arabic document AI.