The next shortlist starts in the archive.

A PDF is a useful record for a person to read, but an awkward unit for repeated comparison. Job titles, training and technical skills can appear in different sections and under different terms. The implemented pipeline ingests text and layout outputs extracted from the CVs, then builds a structured profile for each retained record.

That profile brings together recent titles, skills, professional processes, equipment and qualifications. Original wording stays alongside normalised terms, so a common vocabulary can support search while the source remains available for inspection. The ingestion process also identifies duplicates and records missing inputs, skipped documents and parsing problems. Preparing the pool once gives subsequent searches a reusable foundation.

Alongside each profile, the system builds smaller evidence sections from experience, education, skills and other parts of the CV. It preserves text offsets and, where the extraction output supports them, page and document references. Candidate profiles serve the ranking task; these more precise passages serve the explanation.

A vacancy contains several kinds of information: the role itself, activities to perform, methods or tools to know, qualifications and preferences. The parser separates these into a structured request. Rules and domain dictionaries retain specialist vocabulary and distinguish the strength of the language used in the brief, including requirements described as essential or preferred.

This matters for technical roles. Experience in operating equipment, setting it up and performing a particular process can describe related but different capabilities. The search needs those distinctions to survive the journey from document to query. A broad resemblance between two texts is only part of the evidence for professional fit.

Structured checks add another layer. For example, the implementation can compare stated qualifications and experience requirements with available CV fields. Missing qualification or experience data remains an uncertainty in those checks. The parsed request is saved with the search, making its interpretation inspectable when a result needs investigation.

The retrieval architecture combines two search approaches. Keyword search gives exact professional terms a strong role. Semantic retrieval compares representations of the vacancy and candidate profiles, adding a route to relevant material expressed differently. The two result lists are combined before a narrower group of candidates receives a further ranking pass.

That second pass compares the job description with compact candidate profiles. An additional screening layer records strengths, concerns and requirement gaps. The returned order therefore reflects several signals that can be inspected separately. The system ranks candidates as complete records, with evidence sections attached afterwards, so several passages from one CV do not become several apparent candidates.

The indexed path is implemented in OpenSearch. A deterministic local mode supports development and repeatable checks without loading the inference models. Separating preparation, retrieval and focused ranking is the architectural response to the immediate-shortlist objective: document processing happens ahead of the new search, and deeper comparison concentrates on a selected group.

Method diagram

Close

The final output includes a rank, component scores, matched terms, a brief explanation and selected CV passages for each candidate. The explanation identifies relevant evidence and surfaces recorded concerns or filter gaps. This gives the recruiter a practical starting point for checking the original document and deciding what to clarify in a conversation.

The ten-person shortlist is a requested review format. The number of returned candidates is configurable in the implementation. A result's position expresses the system's ordering for that vacancy; the recruiter still needs to assess current availability, confirm experience and qualifications, and decide how to progress the application.

Each search also retains its parsed brief, search clauses, intermediate candidate lists, evidence references and configuration versions. If a promising profile is missing, or an unsuitable one appears too high, the team can investigate how the request was interpreted and where the ordering changed. That record supports specific improvements to the workflow.

Evaluation tooling compares the individual retrieval approaches with the combined pipeline. Recall measures whether relevant candidates enter the search results; ranking metrics examine where they appear. Applied to recruiter-reviewed vacancies, these measures provide a way to distinguish broader coverage from a better shortlist, and to examine whether an additional ranking stage earns its place.

The prototype establishes a working retrieval core: a vacancy becomes a structured query, ranked candidates and source evidence. Preparing the archive ahead of time is what makes the immediate-shortlist objective practical. Full-scale validation pairs response-time testing across the complete archive with recruiter judgments on the returned matches. The aim is a pool that the agency can search repeatedly, question precisely and improve through review.