RAG for job recommendations: building a semantic matching engine

Almost every professional has had the frustrating experience of setting up an alert on a traditional job board for "Senior Software Engineer" only to receive daily emails promoting entry-level QA jobs, unrelated IT helpdesk positions, or listings located in cities thousands of miles away.

Legacy job platforms rely on simplistic title-string matching. If an employer uses a creative title like "Product Infrastructure Craftsman", traditional alert engines miss it completely. If a job posting mentions "software engineer" in the company boilerplate, the keyword scanner flags it as a match regardless of the actual role.

Modern platforms (like JobsDart) solve this challenge by applying Retrieval-Augmented Generation to candidate recommendations. By treating the candidate’s career history and the employer’s job posting as rich semantic documents, RAG discovers deep alignment that goes far beyond surface-level job titles.

A production job recommendation engine operates as a bidirectional matching system combining vector retrieval with structured relational filters.

• 1. Multi-Vector Candidate Profiling: Instead of averaging an entire resume into a single blurry vector, the engine generates distinct sub-vectors: Core Technical Skills, Management & Leadership Depth, and Domain Experience (e.g. Fintech, Healthcare).
• 2. Structured Constraint Filtering: Hard parameters — minimum desired salary, remote vs hybrid preference, visa sponsorship requirements, and seniority tier — are indexed in PostgreSQL relational columns.
• 3. Two-Stage Retrieval: When a candidate browses their feed (or when a background worker evaluates new daily listings), the engine runs a SQL pre-filter to enforce hard constraints, followed by a pgvector HNSW cosine search against the candidate’s skill vectors.
• 4. RAG Match Explanation: The top 5 matching jobs are passed to a lightweight LLM that generates a personalized match rationale: "We recommend this role because their distributed streaming stack matches your 4 years of Kafka experience at your previous company."

A major challenge in traditional collaborative filtering recommendation systems (like matrix factorization) is the cold-start problem: a new user has zero click history, so the algorithm cannot predict what they want.

RAG eliminates the cold-start problem completely. When a user creates an account and uploads their resume, the ingestion worker parses their career history, generates their semantic skill vectors, and immediately queries the active job catalog.

Within 500 milliseconds of signing up, the candidate receives a curated feed of job opportunities that closely match their real-world experience, without requiring them to fill out tedious 30-step preference questionnaires.

To scale a job recommendation engine to millions of candidates and tens of thousands of daily job listings, engineering teams balance real-time user browsing with asynchronous push alerts.

When a user visits their dashboard, execute real-time vector queries against the active job index using PostgreSQL pgvector with HNSW indexing to deliver instantaneous sub-50ms feed rendering.

When an employer posts a new job, publish an event to a Kafka or RabbitMQ queue. An asynchronous worker queries candidate vectors that match the job criteria, bundling the best candidates into daily notification digests for recruiters.

• Use HNSW vector indexes in PostgreSQL for sub-50ms candidate-to-job similarity queries
• Enforce strict SQL pre-filtering on salary and location before running vector distance calculations
• Generate transparent match explanations so candidates understand why a job was recommended
• Decouple real-time dashboard searches from asynchronous background email push notifications