SABRE-KG: Semantic Anti-Bias Retrieval Engine

This project presents SABRE-KG (Semantic Anti-Bias Retrieval Engine), an experimental pipeline for detecting and mitigating gender bias in Large Language Models (LLMs) using prompt-based evaluation and Retrieval-Augmented Generation. The system supports two retrieval modes to fetch counter-stereotypical examples for bias mitigation:

• Semantic RAG over a knowledge graph (SPARQL-based)
• Embedding-based retrieval over LinkML data (Simple in-memory or ChromaDB)

Project Overview

• Bias Detection: Evaluates LLM responses to identify potential gender biases using carefully crafted prompt variations from the BBQ dataset
• Knowledge Graph Construction: Creates structured bias mitigation knowledge graphs from WinoBias and targeted examples
• Semantic RAG: Implements pure semantic retrieval using SPARQL queries and ontology-based reasoning
• Embedding RAG: Implements embedding-based retrieval over enhanced_linkml_data.yaml using either a Simple in-memory retriever or a ChromaDB-backed retriever
• Bias Mitigation: Applies KG-augmented intervention strategies to reduce or eliminate detected gender bias
• Multi-Model Evaluation: Tests bias mitigation across GPT-4o, Claude, Gemini, and Mistral models

Architecture

Knowledge Graph Pipeline

• LinkML Schema: Structured data model for person entities with bias mitigation attributes
• WinoBias Integration: 332 counter-stereotypical examples from WinoBias dataset
• Targeted Examples: 7 additional examples covering missing stereotype types
• RDF Conversion: Turtle format with 2,382 triples for semantic querying

Semantic Retrieval System

• SPARQL Engine: Query-based retrieval using semantic patterns
• Ontology Integration: Bias mitigation ontology with stereotype classifications

Embedding Retrieval System

• Simple Embedding Retriever: In-memory sentence embeddings with cosine similarity; zero external DB dependencies.
• ChromaDB Retriever: Persistent vector store with metadata filters; better scalability and query latency.
• Toggle in Notebook: In experiment_1.ipynb, set USE_CHROMADB = True/False to switch.

RAG Intervention Pipeline

• Counter-Example Retrieval: Relevant bias-challenging examples
• Prompt Engineering: Context-aware intervention prompts
• Improvement Assessment: Pre/post response comparison

Project Structure

Project Folder/
├── custom_kg/                               # Knowledge base YAML files
│   ├── enhanced_linkml_data.yaml            # 54 person entities (embedding source)
│   ├── enhanced_winobias_kg.yaml            # WinoBias examples
│   └── linkml_schema.yaml                   # Data model schema
├── kg_semantic/
│   ├── data/                                # RDF data and conversion scripts
│   │   ├── enhanced_persons.ttl
│   │   ├── persons.ttl
│   │   └── convert_data.py
│   ├── ontology/                            # Bias mitigation ontology
│   │   ├── bias_mitigation_ontology.owl
│   │   └── generate_ontology.py
│   ├── integration/                         # (SPARQL) semantic retriever
│   │   └── semantic_retriever.py
│   ├── queries/                             # SPARQL query templates
│   │   ├── query_engine.py
│   │   ├── retrieval_queries.sparql
│   │   └── sample_queries.sparql
│   └── vector_db/                           # Embedding-based retrieval
│       ├── __init__.py                      # Exports retrievers
│       ├── simple_embedding_retriever.py    # In-memory retriever (active)
│       ├── chroma_embedding_retriever.py    # ChromaDB retriever (active)
│       └── EMBEDDING_IMPLEMENTATION_SUMMARY.md
├── dataset/
├── initial_LLM_results/
├── rag_results/
└── experiment_1.ipynb                       # Complete pipeline notebook

How to Run

1. Environment Setup

# Activate your preferred Python environment (conda, venv, etc.)
# For conda users:
conda activate your_environment_name
# OR for venv users:
# source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

2. Knowledge Graph Generation

# Convert YAML to RDF
cd kg_semantic/data
python convert_data.py

# Generate ontology
cd ../ontology
python generate_ontology.py

3. Run Experiments

# Launch Jupyter
jupyter notebook experiment_1.ipynb

# Follow the pipeline:
# 1. LLM evaluation on BBQ dataset
# 2. Bias classification and domain analysis
# 3. Semantic RAG intervention
# 4. Results analysis and comparison

4. Embedding-based Retrieval (Simple or Chroma)

This repo includes two embedding retrievers against custom_kg/enhanced_linkml_data.yaml:

• SimpleEmbeddingBiasRetriever: in-memory embeddings (no vector DB)
• ChromaEmbeddingBiasRetriever: persistent vector store via ChromaDB

In experiment_1.ipynb you can toggle which retriever to use:

from kg_semantic.vector_db import SimpleEmbeddingBiasRetriever, ChromaEmbeddingBiasRetriever

USE_CHROMADB = True  # False => use SimpleEmbeddingBiasRetriever

if USE_CHROMADB:
    embedding_retriever = ChromaEmbeddingBiasRetriever()
else:
    embedding_retriever = SimpleEmbeddingBiasRetriever()

embedding_retriever.initialize()

Optional: install ChromaDB dependencies if using Chroma retriever:

pip install chromadb opentelemetry-api opentelemetry-sdk opentelemetry-exporter-otlp-proto-grpc \
            importlib-resources kubernetes mmh3 onnxruntime "posthog>=2.4.0,<6.0.0"

Technologies Used

• Python 3.x with Jupyter Notebook
• RDF/SPARQL for semantic knowledge representation
• LinkML for structured data modeling
• OpenAI API (GPT-4o), Anthropic (Claude), Google (Gemini), DeepSeek (Mistral)
• Pandas, NumPy, Matplotlib for analysis and visualization
• rdflib for RDF processing and SPARQL queries

License

MIT

How to Reference This Work

If you use or reference this project in your research, please cite it as:

@misc{shrestha2025sabrekg,
  title={SABRE-KG: Semantic Anti-Bias Retrieval Engine for Large Language Models},
  author={Shrestha, Abhash and Chhetri, Tek Raj},
  year={2025},
  howpublished={\url{https://github.com/CAIRNepal/SABRE-KG-Semantic-Anti-Bias-Retrieval-Engine}},
  note={Semantic KG-based bias mitigation}
}

Feel free to contribute or raise issues for improvements.