if you checked tehran data rag project : The goal of building this agentic workflow is to identify and track logical fallacies in Persian tweets. The architecture of this system is based on HyDE (Hypothetical Document Embeddings) RAG and LLM-as-a-judge.

The first agent attempts to detect and hypothesize the fallacy. Meanwhile, the judge agent (the second agent) utilizes a RAG system containing the description, name, and a few examples of each fallacy. The first agent must use the HyDE mechanism to convince the judge agent that its hypothesis is correct. If the explanation generated by the first agent falls within the top three most relevant retrieved results, the judge model is convinced.

Leveraging agentic AI and RAG eliminates model hallucinations and bypasses the need for heavy, resource-intensive models. The Persian language has a complex structure where logical concepts are often expressed through irony, metaphors, and unconventional phrasing; because of this, heavy models on their own struggle to conduct proper logical analysis. Furthermore, available datasets in the Persian language are extremely limited.