GCC-RAG-HardCases

A deterministic synthetic benchmark for evaluating difficult Retrieval-Augmented Generation (RAG) scenarios across Arabic and English knowledge contexts.

Overview

720 validated synthetic cases

• 8 RAG failure families
• 6 GCC country contexts
• Arabic, English and mixed-language scenarios
• deterministic ground truth
• independent validation
• no real customer data
• no PII
• no copied documents

Why this benchmark exists

Enterprise RAG systems often perform well on simple lookup tasks while failing on harder knowledge-base situations.

This benchmark focuses on:

• stale document versions
• conflicting source authority
• contradictory numerical values
• similar entity names
• questions with no supported answer
• multi-document reasoning
• Arabic/English cross-lingual retrieval
• date-dependent policies

Failure families

stale_version

The system must select the latest applicable version rather than outdated information.

authority_conflict

Two sources disagree. The explicitly higher-authority source is correct.

numeric_conflict

Documents contain contradictory numeric values.

entity_ambiguity

Similar entity names must not be confused.

no_answer

The requested information is absent. Correct behavior is abstention rather than hallucination.

multi_hop

The answer requires joining evidence across multiple documents.

cross_lingual

The query and supporting evidence are in different languages.

date_boundary

The answer depends on which policy was effective at a specific date.

Coverage

{
  "cross_lingual": 90,
  "stale_version": 90,
  "date_boundary": 90,
  "authority_conflict": 90,
  "entity_ambiguity": 90,
  "numeric_conflict": 90,
  "no_answer": 90,
  "multi_hop": 90
}
Country contexts:
{
  "SA": 120,
  "BH": 120,
  "KW": 120,
  "QA": 120,
  "AE": 120,
  "OM": 120
}

Language modes:
{
  "ar": 240,
  "mixed": 240,
  "en": 240
}

Example structure
{
  "query": "What is the current approval threshold?",
  "documents": [
    {
      "doc_id": "DOC-A",
      "language": "en",
      "version": 2,
      "effective_date": "2026-02-01",
      "authority_rank": 2,
      "source_kind": "manual",
      "content": "The approval threshold is 300 GCCU."
    },
    {
      "doc_id": "DOC-B",
      "language": "en",
      "version": 3,
      "effective_date": "2026-08-01",
      "authority_rank": 2,
      "source_kind": "manual",
      "content": "The approval threshold is 325 GCCU."
    }
  ],
  "expected": {
    "answerable": true,
    "answer": "325 GCCU",
    "supporting_doc_ids": ["DOC-B"],
    "rejected_doc_ids": ["DOC-A"],
    "rationale_code": "numeric_latest"
  }
}

Validation
The complete release passed deterministic validation:
- 720 / 720 PASS
- 720 unique IDs
- 720 unique fingerprints
- schema validation
- provenance validation
- document-reference validation
- family-specific oracle checks
The validator is included in this repository.
Ground truth
Expected answers are not generated by an LLM.
Scenario rules determine the correct:
- answer
- supporting documents
- rejected documents
- answerability status
- rationale type
This keeps evaluation reproducible.
Arabic scope
Arabic content is synthetic Modern Standard Arabic-style benchmark text.
The dataset does not claim:
- Saudi dialect authenticity
- Emirati dialect authenticity
- Kuwaiti dialect authenticity
- Qatari dialect authenticity
- Bahraini dialect authenticity
- Omani dialect authenticity
Synthetic provenance
No real:
- customer knowledge bases
- internal corporate documents
- personal data
- government documents
- confidential policies
are included.
All examples are synthetic.
Intended use
Useful for evaluating:
- enterprise RAG
- multilingual RAG
- hallucination resistance
- retrieval pipelines
- citation systems
- AI agents using knowledge bases
- document migration
- knowledge-base regression testing
Commercial customization
Need a private benchmark for your RAG system?
We can generate a custom benchmark adapted to:
- your knowledge-base structure
- your document taxonomy
- your languages
- policy hierarchy
- document versions
- private entity structures
- temporal rules
- known production failure modes
- your expected JSON schema
Typical delivery:
Your KB structure ? synthetic benchmark ? deterministic ground truth ? validator ? QA report
The commercial value is not simply producing more rows.
It is reproducing the failure modes that matter to your actual system.
Limitations
This release does not measure:
- OCR
- speech
- embedding quality alone
- real-world document distributions
- production latency
- regulatory correctness
- Gulf dialect authenticity
Reproducibility
Generation is deterministic and seed-based.
Default seed:
20260930
Generator and validator source are provided.
License
MIT.

---

## RegalFire — Custom / Private Dataset Work

RegalFire builds custom AI datasets, evaluation sets and data pipelines for:

- AI agents
- computer-use systems
- multimodal models
- world models
- RAG systems
- code agents
- enterprise AI

Available services include:

- synthetic data generation
- private evaluation datasets
- agent trajectories
- failure / recovery datasets
- multimodal RGB / segmentation / state-action data
- web data acquisition
- cleaning and deduplication
- structured dataset packaging
- continuous dataset production

For custom or private work:

**Email: ootiris@gmail.com**

Hugging Face: **RegalFire**