AI Parsing Guidelines

This document explains how AI systems (LLMs, extraction pipelines, or parsing engines) should transform unstructured CV data into the Barba-CV JSON format.

Barba-CV is designed to act as the deterministic structural layer between probabilistic AI extraction and structured HR datasets.

This page focuses on extraction and field-level mapping behavior, while LLM Integration focuses on prompt packaging, orchestration, and schema-guided workflow design.

1. Goal of AI parsing

The goal of an AI parser is to convert:

Unstructured CV (PDF / DOCX / HTML / text)
        ↓
AI extraction
        ↓
Barba-CV JSON structure

The parser should map information into the Barba-CV structure without inventing data.

Missing information should remain empty.

2. Never invent information

If a field cannot be extracted with confidence, it should be left empty.

Examples:

"date_of_birth": ""
"phone": ""
"organization": ""

AI systems must not hallucinate information.

3. Prefer structured extraction over free text

Whenever possible, the AI should structure information into the correct fields.

Example:

Instead of:

"profile_summary": "Worked at ACME as a software engineer from 2020 to 2023"

Use:

"experiences": [{
  "organization": "ACME",
  "role_title": "Software Engineer",
  "start_date": "2020",
  "end_date": "2023"
}]

4. Handling incomplete dates

Dates in CVs are highly variable.

Allowed examples:

"2020"
"Jan 2021"
"2021-2023"
"September 2019"
"Present"

Do not attempt to normalize aggressively.

Dates should remain human readable.

5. Experience extraction

Each professional role should become one element in experiences.

Example:

"experiences": [
  {
    "organization": "Example Company",
    "role_title": "Senior Engineer",
    "start_date": "2020",
    "end_date": "2023",
    "tasks": [],
    "achievements": []
  }
]

Descriptions should be split when possible:

  • responsibilities → tasks
  • measurable outcomes → achievements

6. Education extraction

Each education entry becomes one element in education.

Fields to extract when available:

  • school
  • degree
  • field
  • dates
  • location

7. Skills classification

Skills should be categorized when possible:

"skills": {
  "it_skills": [],
  "hard_skills": [],
  "soft_skills": []
}

Guidelines:

Category Meaning
it_skills programming, tools, software
hard_skills professional capabilities
soft_skills interpersonal or behavioral skills

If classification is unclear, place the skill in hard_skills.

8. Position sought

The field position_sought describes the candidate’s professional target.

Examples:

"position_sought": [
  "Full Stack Developer",
  "Python Expert"
]

If the CV includes a headline or title, it should be mapped here.

9. Languages

Languages should include:

{
  "language": "English",
  "level": "Fluent"
}

Levels remain free text to allow different conventions.

10. Certifications

Certifications should include issuer and dates when available.

Example:

{
  "name": "AWS Certified Solutions Architect",
  "issuer": "Amazon",
  "date_obtained": "2022"
}

11. Project achievements

Major projects or consulting missions can be extracted into project_achievements.

Example:

{
  "title": "ERP Implementation",
  "client": "Manufacturing Group",
  "role": "Project Manager",
  "period": "2021-2022"
}

12. Metadata population

The meta block contains operational information about parsing.

Typical fields include:

"meta": {
  "cv_uuid": "",
  "processor_engine": "",
  "ats_processed": false
}

AI systems should populate metadata only when known.

13. Extensions

Custom system information must be placed inside extensions.

Example:

"extensions": {
  "ats_id": "12345",
  "client_reference": "ABC"
}

This prevents breaking the core schema.

14. Output validation

After extraction, the generated JSON should:

  1. validate against barba-cv.schema.json
  2. contain only supported fields
  3. respect the root CV structure

Summary

AI parsing with Barba-CV follows a simple rule:

Extract faithfully. Structure deterministically. Never invent data.

This ensures reliable CV parsing while keeping the schema compatible with real-world documents.