12. Conclusion

The evaluated local Gemma 4 extraction pipeline is technically viable and factually conservative, but it is not a complete representation of the source CV. Across 24 predefined occupational categories and 2,484 PDF CVs, weighted precision is 94.8%, weighted recall is 65.5%, weighted F1 is 77.4%, and completeness is 68.2%. The system is more likely to omit evidence than to invent it. This makes the output useful for positive-evidence indexing but unsafe as the sole source of negative evidence or automatic exclusion.

The strongest fields are Experience, Profile, and Skills, although even Skills requires source fallback. Education, Certifications, and Languages can support retrieval but cannot safely support absence-based filters. Document metadata, category classification, normalized fields, and seniority should be excluded from authoritative filtering or recomputed from source evidence.

Reasoning-disabled inference is defensible for a constrained extractive task because it can reduce latency, output variance, and leakage of internal reasoning. However, it may also reduce recall on ambiguous and compositional cases. The current data do not establish causality; a controlled on-versus-off benchmark is required. The prompt contributes to high precision through strict verbatim rules, but it needs a completeness pass, provenance-bearing output, constrained decoding, and separation of evidence from normalization.

The recommended deployment is a field-aware hybrid RAG system. It should index both validated structured data and original section-level text, decompose job descriptions into typed criteria, compare each criterion with compatible evidentiary roles rather than rigidly exclusive fields, execute raw-text fallback for missing fields, and display source excerpts for every material match. Human review must remain responsible for candidate shortlisting and decisions. Under these controls, the extracted data can materially accelerate CV search while preserving the source document as the final authority.

Deployment decision: Proceed with controlled, evidence-first RAG integration. Do not deploy the extracted JSON as a stand-alone candidate database, a complete representation of each CV, or the sole basis for candidate exclusion.