Problem
Keyword search misses candidates whose experience is relevant but phrased differently. A backend matching service needs semantic retrieval while still enforcing tenant isolation, filters, and explainable ranking signals.
Professional Case Study · AI Search
Semantic Candidate-Job Matching
A company-owned matching system that converts resumes and job descriptions into embeddings, stores searchable candidate vectors, applies organization-safe metadata filters, and ranks candidates by semantic fit.
Confidentiality: Source code and internal implementation details are private. This case study describes my contribution and general engineering decisions without disclosing proprietary information.
Candidate resume → text extraction → embedding
↓
Candidate vector stored with tenant, skills, seniority, and location metadata
↓
Job description → embedding
↓
Similarity search + metadata filters → ranked candidates
Key Decisions
- Store vectors with metadata so semantic search and business filters run together.
- Keep tenant or organization identifiers mandatory in every query path.
- Rank by vector similarity first, then adjust with structured signals such as skill overlap and recency.
- Return reason snippets so results are easier to inspect than a raw score.
Failure Handling
- Embedding generation failures are queued for retry without losing source records.
- Records with missing or stale vectors are excluded from semantic ranking until repaired.
- Filters are applied before ranking to prevent cross-organization leakage.
- API responses include empty-result states rather than falling back to unsafe broad searches.
Trade-offs
- Embeddings improve recall, but exact filters are still required for recruiter workflows.
- Similarity scores are useful for ordering, not as absolute truth.
- Precomputing candidate vectors adds ingestion work but keeps query latency predictable.
Results
The system applies embeddings, vector storage, metadata constraints, ranking, and guardrails to practical recruiter workflows.