How one team rebuilt broken job search using PostgreSQL and hybrid retrieval

A careers page search tool was producing poor results — returning JavaScript roles for 'Java' queries and failing entirely on misspellings — because it relied on simple keyword matching against concatenated job metadata. The engineering team investigated the issue after spotting skipped end-to-end tests, though they carefully distinguished confirmed production bugs from unverified diagnoses before deciding on a fix. Rather than adopting a dedicated search engine like Elasticsearch, they chose to build a hybrid retrieval system within their existing Next.js and PostgreSQL stack, using OpenAI embeddings already integrated into their platform. The new design combines weighted lexical search with semantic matching, allowing it to handle spelling variations, technology aliases, and intent-based queries while keeping company boundaries and explicit filters intact. The team acknowledged this leaner architectural choice still required significant engineering effort, including defining what a genuinely useful result set looks like under real-world conditions such as stale data and provider failures.