Contextual Compression of Unstructured Resumes using VectorStore-Based Retrieval Augmented Generation

Extracting structured data from unstructured resumes and CVs is an intricate & extremely difficult task and it is also prone to mistakes especially during the Application Tracking System (ATS) recruitment where speed & precision are of the utmost importance. This paper introduces a Contextual Compression System intended for resume data extraction as well as disorganized content to be transformed into a structured JSON format. The system leverages machine learning models, OpenAI embeddings, & the retrieval-augmented generation (RAG) techniques. It incorporates advanced NLP methods, such as FAISS-based vector storage, LangChain for document loading, & a modular Flask API for seamless integration. Key tools include a Contextual Compression Retriever & OpenAI-powered LLMs for enhanced semantic understanding. The system was evaluated on 100 resumes using metrics like Cosine Similarity (CS: 0.9296 ± 5.1%), Pearson Correlation (PC: 0.0821 ± 0.12%), & Spearman Correlation (SC: 0.0722 ± 0.12%). The highest accuracy was achieved in the Education (CS: 0.960 ± 0.035) & Work Experience (CS: 0.940 ± 0.045) sections. By enabling structured data extraction across essential resume sections, the system improves recruitment analytics, supporting HR teams in making data-driven decisions with enhanced efficiency & precision.