Attempting to automate recruitment with pure vector search routinely hits a wall: the system eagerly suggests a strong senior developer for an enterprise architect role simply because their resumes share the same buzzwords. Embeddings excel at catching abstract textual similarity, but they are utterly blind to true seniority, project scale, and actual candidate responsibilities. In a benchmark test run by an IT recruitment platform across 476 resume-vacancy pairs, pure vector search delivered an underwhelming Top-10 precision of just 43.7%.
The solution lies in moving to a two-stage pipeline powered by an LLM-as-a-judge. Engineers restructured the ranking workflow: an initial vector filter discards obvious non-matches, while a language model reranks the candidates against a rigid, structured checklist. The LLM evaluates the candidate based on their actual role, scope of responsibility, and tech stack, returning strict JSON. The core safeguard against hallucinations is mandatory citation validation: the model must back up every requirement match with a direct resume quote up to 200 characters. No verbatim quote means the match score drops instantly.
This two-stage approach pushed strict Top-10 precision to 66.3%. More importantly for unit economics, benchmark tests across four models showed that the compact `gpt-4.1-mini` matched the quality of the flagship `gpt-4o` while slashing inference costs by 80%. Costly reasoning models are simply unnecessary here: effective enterprise scoring relies on strict structural constraints and automated fact verification, not on burning budget on oversized context windows.