LLM MOE models for RAG search

Introduction

RAG for CV search and the same-context requirement.

1. Introduction

Retrieval-augmented generation (RAG) couples an information-retrieval subsystem with a generative model so that answers and decisions can be conditioned on external evidence rather than only on model parameters [1]. In a CV-search system, the retrieved evidence is not general encyclopedic text; it is candidate-specific evidence such as employment history, skills, education, certifications, languages, and source passages. The operational objective is therefore not merely semantic similarity. It is evidence retrieval under a typed schema, followed by criterion-level explanation.

A defensible CV RAG system separates at least three computational stages. First, the source CV is converted into structured fields and source-aligned text spans. Second, the job description is decomposed into atomic criteria, for example a required programming language, minimum experience duration, degree level, location constraint, or certification. Third, each criterion is compared only with semantically corresponding CV evidence. The generative model may summarize or explain the retrieved evidence, but it should not replace the evidence or infer unobserved credentials.

The distinction between positive evidence and negative evidence is essential. A structured value that is present and supported can be used as positive evidence. A structured value that is absent cannot automatically be used as negative evidence because extraction recall is materially below precision. The original document remains authoritative and must remain retrievable.

1.2 The same-context requirement

The same-context requirement is the central design constraint for CV retrieval. A skill requirement in a job description should be matched against skill statements and professional-experience evidence, not against hobbies, personal interests, file metadata, or unrelated course titles. An education requirement should be matched against degree and institution evidence, not employer training unless the source explicitly establishes an equivalent credential. A language requirement should be matched against language proficiency evidence, not nationality, residence, or inferred name origin.

Formally, the retrieval process should preserve the evidentiary role of both the query criterion and candidate evidence. This does not require a rigid one-fact-one-field ontology: a technology named in an employment bullet can be valid evidence for both experience and skill, and a leadership accomplishment can support both responsibility and competency. Cross-field semantic similarity is acceptable when the source statement has a compatible evidentiary role, but unrelated contexts such as hobbies, metadata, or generic interests must carry a strong penalty and cannot by themselves satisfy mandatory criteria. This prevents semantic leakage while allowing multi-role evidence.

Figure 1. Field-aligned extraction and RAG architecture.
Figure 1. Field-aligned extraction and RAG architecture.

Source: pooled PDF-grounded evaluation workbooks; calculations in the combined analysis.

Interpretation. The architecture separates offline extraction from online retrieval. Both the job criterion and the CV evidence are typed before matching. Structured retrieval is supplemented by source-section fallback so that omitted JSON fields do not become false negatives.

Limitation. The diagram is a recommended architecture derived from the evaluation results; it is not a direct measurement of the supplied implementation.

Open SVG

1.3 Research questions and contributions

Table 1. Research questions and corresponding evidence.
ID Research question Evidence used
RQ1 How accurately and completely does the local model reproduce PDF-supported CV facts? Atomic-fact precision, recall, F1, completeness, and unsupported-content rate.
RQ2 Which occupational datasets and CV segments are strongest or weakest? Dataset and segment metrics, heatmaps, outcome distributions, and error logs.
RQ3 Is the extracted JSON suitable for RAG-based CV search? Field reliability, missingness risk, source fallback requirements, and ranking implications.
RQ4 What are the performance implications of the supplied runtime and reasoning-disabled configuration? Analytical latency model, concurrency settings, MoE/QAT characteristics, and reasoning-mode trade-offs.
RQ5 How does the prompt design shape the observed precision-recall profile? Prompt-contract analysis linked to omissions, segmentation errors, and normalization failures.

The report contributes a pooled empirical baseline across heterogeneous occupational datasets, a formal definition of the metrics used in the attached reports, a high-level systems analysis of the local inference configuration, and a deployment-oriented RAG design that explicitly accounts for extraction uncertainty.