کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Effهcient Dual-Retrieval Graph-based RAG for Persian Technical Documents in an Offline Oil Refinery Setting

نویسندگان :

Zohreh Ansari¹ Mansoor Khademi² Gholamreza Bagheri Di³

1- شرکت پالایش نفت اصفهان 2- شرکت پالایش نفت اصفهان 3- شرکت پالایش نفت اصفهان

کلمات کلیدی :

Retrieval Augmented Generation،LLM،Knowledge Graphs،Graph-based RAG

چکیده :

Retrieval Augmented Generation (RAG) models enhance the performance of Large Language Models (LLMs) in more accurate and contextually relevant responses by integrating them with external knowledge of documents. Access to reliable, domain-specific knowledge in Persian as a low-resource language remains a practical challenge in industrial settings, especially when internet connectivity, data governance, or confidentiality restricts the use of cloud services. This paper presents an offline retrieval-augmented generation algorithm designed for Persian refinery operation manuals and books. The proposed approach combines semantic vector retrieval with a lightweight knowledge-graph index derived from the same corpus, enabling both precise (entity-anchored) retrieval and broader (topic-level) context discovery. The system runs end-to-end on local Esfahan Oil Refinery infrastructure using the optimized offline multilingual LLM, and multilingual embedding model with long-context retrieval on Persian language. Experimental results verify that the system achieved high perceived correctness and usefulness in comparison with the other well-known offline and online LLM models