AI Resume Screening System Based on Semantic Embedding and FAISS Vector Search

This project presents an AI resume screening system that integrates semantic similarity calculation and structured skill matching, aiming to address the limitations of traditional keyword matching. The system uses Sentence Transformers to generate embedding vectors, FAISS for fast vector retrieval, spaCy for skill extraction, and ranks candidates intelligently via a hybrid scoring model to improve the efficiency and accuracy of recruitment screening.

Traditional resume screening systems rely on keyword matching, which has limitations such as inability to recognize synonyms, ignoring contextual semantics, and vulnerability to keyword stuffing. With the advancement of NLP technology, semantic understanding-based screening has become feasible. This project provides a complete AI-driven solution to achieve intelligent evaluation through vector embedding and semantic similarity calculation.

The system adopts a modular design, with core components including:

• Embedding Generation Layer: Sentence Transformers convert text into high-dimensional vectors
• Vector Storage Layer: FAISS implements efficient approximate nearest neighbor search
• Skill Extraction Layer: spaCy performs named entity recognition and skill extraction
• Scoring Fusion Layer: Combines semantic similarity and skill matching degree to calculate rankings
• Interactive Interface Layer: Streamlit builds web dashboards
• Data Persistence Layer: SQLite stores evaluation results

Semantic Similarity Calculation

Pre-trained sentence embedding models are used to convert job descriptions and resumes into dense vectors, capturing semantic relationships (e.g., the vector distance between "Python development" and "Python programming" is close). Cosine similarity is used for measurement (normalized to [0,1]).

FAISS Vector Retrieval Optimization

IVF indexing is used to divide the vector space into cluster centers. When querying, only relevant clusters are searched, reducing the computational complexity of large-scale resume databases.

Hybrid Scoring Model

Final score = 0.7 × semantic similarity + 0.3 × skill overlap, balancing overall semantic fit and precise skill matching.

Skill Extraction and Matching

spaCy is used to extract structured information such as technical skills. Fuzzy matching via a skill dictionary and word form normalization are applied to handle variant expressions (e.g., ReactJS and React.js are recognized uniformly).

The system is suitable for enterprise recruitment. Its core technology can also be extended to:

• Internal talent pool search: Quickly locate employees with specific experience
• Project personnel matching: Recommend suitable team members
• Career development advice: Analyze gaps between resumes and target positions
• Academic literature recommendation: Personalized paper report recommendations

Current limitations: FAISS index in-memory storage loses data on restart; lack of in-depth resume format processing; limited multilingual support. Improvement directions: Introduce persistent vector databases (Milvus/Pinecone); integrate stronger document parsing engines (Unstructured); support multilingual embedding models (mBERT/XLM-R).

This project demonstrates the path of transforming cutting-edge NLP technology into practical business tools. Vector search and semantic understanding are reshaping the field of information retrieval (recruitment, e-commerce, knowledge management, etc.). The AI resume screening system improves matching quality and efficiency; in the future, it will be more accurate and user-friendly, changing the way talent is discovered and evaluated. Developers need to master tools like Sentence Transformers and FAISS to build intelligent applications.