Μεγάλα γλωσσικά μοντέλα με χρήση RAG για την αντιστοίχιση υποψηφίων με θέσεις εργασίας
RAG-based LLMs for job matching tasks
View/
Keywords
Αξιολόγηση βιογραφικών ; Κατάταξη υποψηφίων ; Περιγραφή θέσης εργασίας ; Διανυσματική βάση δεδομένων ; Τοπικό γλωσσικό μοντέλο ; Αναγνώριση κειμένουAbstract
This thesis presents the design and implementation of an innovative automated system for curriculum vitae (CV) evaluation and candidate ranking using Artificial Intelligence (AI) technologies. The primary objective of the application is to automate the recruitment process by measuring the degree of relevance between candidates’ qualifications and the requirements of a specific job position, as defined in a Job Description.
The system was developed in Python using the Streamlit framework and supports multimodal data input, enabling the processing of PDF documents, image files (JPG/PNG), and bulk ZIP archives. To extract textual information from visual documents, the EasyOCR Optical Character Recognition (OCR) technology was integrated into the platform.
The system architecture is based on the advanced Retrieval-Augmented Generation (RAG) paradigm, which enables the Microsoft Phi-3 Large Language Model (LLM) to dynamically analyze and interpret data in real time. Information retrieval is performed through the ChromaDB vector database, where CVs are transformed into mathematical vector representations (embeddings) to facilitate semantic search and similarity matching. The orchestration and integration of all system components are achieved through the LangChain framework.
The final user interface provides HR professionals with a dual-view comparison environment. On the left side, newly submitted candidates are ranked according to their suitability for the target position, while on the right side, the most relevant candidates are retrieved from the pre-existing vector database, which is built through the ingestion process from a public dataset of 962 CVs. Candidate ranking is performed automatically in descending order based on a suitability score ranging from 0 to 100%.
In addition, the system generates a qualitative summary for each candidate and proposes three tailored interview questions, focusing on identified skill gaps, missing qualifications, or weaknesses within the candidate’s profile. By leveraging local model execution through the Ollama platform, the proposed solution ensures full privacy, security, and confidentiality of sensitive personal data, eliminating the need for external cloud-based processing.