How to Build an LLM-Powered Smart Resume Screening System

Background and Challenges

Traditional HR resume screening relies heavily on manual review, resulting in low efficiency, lack of intelligence, difficulty quantifying match quality, no natural‑language interaction, and the risk of bias.

System Goals

Natural‑language interaction for HR queries.

Intelligent extraction of structured information from resumes and job descriptions.

Quantifiable matching algorithms that evaluate candidates across multiple dimensions.

Solution Design

Limitations of Classic RAG

Semantic over‑matching leads to insufficient precision.

Cannot handle exact rules such as "work years >= 3" or salary ranges.

Lack of interpretability for similarity scores.

Noise from irrelevant resume sections degrades matching.

Upgraded Approach

The upgraded design merges vector retrieval with structured rule matching, combines hard‑condition filtering with soft‑evaluation, and introduces a multi‑dimensional weighted scoring algorithm.

Architecture Overview

Core Modules

Document Parsing & Information Extraction : Use a multimodal LLM (e.g., LlamaParse) to convert PDF resumes to Markdown and extract key fields such as skills, experience, education, etc.

Vector Index : Embed parsed resumes and store embeddings with metadata in a vector database (e.g., ChromaDB).

Query Understanding : LLM parses HR natural‑language requests or JD files into structured conditions (required skills, work‑year range, salary range, location, etc.).

Semantic Retrieval (Initial Screening) : Retrieve the top‑K most similar resumes based on embedding similarity.

Hard‑Condition Filtering : Apply exact rules (e.g., work_years >= X, required skills present, salary within range) to discard non‑compliant candidates.

Comprehensive Scoring : Compute weighted scores across six dimensions—industry match, skill match, salary match, education match, location match, and personal‑tag match—and combine with the semantic similarity score.

Multi‑Stage Filtering Process

Semantic vector retrieval to create an initial candidate pool.

Hard‑condition filter to enforce precise requirements.

Multi‑dimensional scoring and ranking, with explanations for each dimension.

Scoring Formula

metadata_score = (
    domain_score * 0.35 +   # industry match
    skills_score * 0.35 +   # skill match
    salary_score * 0.10 +   # salary match
    education_score * 0.10 + # education match
    location_score * 0.05 +  # location match
    tags_score * 0.05       # personal‑tag match
)

total_score = semantic_score * 0.2 + metadata_score * 0.8

Result Presentation

Generate a candidate card that displays basic information, detailed match analysis, and an LLM‑generated overall evaluation, allowing HR to quickly assess strengths and weaknesses.

Testing and Demo

A demo with dozens of heterogeneous resumes shows two selection modes: natural‑language query and JD‑file query. The system displays progressive filtering stages, final ranked candidates, and explanatory details for each match.

Future Work

Improve parsing accuracy using knowledge graphs for better skill recognition.

Enhance semantic recall with rerank models.

Refine the multi‑dimensional scoring algorithm, adding priority weighting for preferred attributes.

Extend the pipeline to automate interview invitation generation, schedule suggestions, background checks, and decision‑support analytics.

Build an AI recruitment ecosystem with specialized agents for interview question generation, salary negotiation, and end‑to‑end automation.