Problem
Hays, a global recruitment company, earns by matching candidates to roles. Recruiters search a huge pool of messy, multilingual CVs where one skill appears as "Python", "Python 3" or "Py". The starting search took 20 seconds. Slow or untrusted results cost recruiter time and shortlist quality. In blue-collar hiring, where lead times can be under 24 hours, they cost the placement itself. Hiring is also regulated (high-risk AI under the EU AI Act), so matches must be explainable and free of bias.
Example from the article: a data engineer's CV may say "Python programming" and never "Python", so keyword search misses it.
Solution in brief
Normalize every CV into ontology-anchored concepts early, split it into typed chunks (skills, experience, qualification, all) and embed each type separately. At query time fuse keyword and vector search with reciprocal rank fusion and combine per-type scores with weights the recruiter can tune. Extraction is deterministic first, and bias is checked at every step.
Questions this case answers
- How to rank candidates for a requirement from messy multilingual free-text profiles by normalizing terms into a standard ontology without silently dropping unknown terms
- How to combine keyword and vector search with tunable and explainable per-field weights while keeping the ranking fair and auditable
- How to build a candidate-to-job matching engine over messy multilingual CVs
- How to combine keyword and vector search with tunable, explainable scoring
Tools and methods
- ESCO-anchored four-stage resolver — alias tables, ESCO lookup, fuzzy/embedding match, human-reviewed unknown queue.
- Typed-chunk embeddings — one vector space per chunk type.
- Hybrid retrieval with reciprocal rank fusion — keyword plus semantic search.
- Weighted per-type scoring — coverage multiplier and weight renormalization.
- Eval pipeline with bias judge — precision, recall, F1 per CV.
- Databricks Vector Search, Lakebase, Unity Catalog, MLflow, Databricks Apps, Genie.
Google or a one-time email code. No password.