How dt-service-consulting/unbiasai works
Project Type and Purpose
• This is an open-source Python package (“UnBiasAI”) designed to analyze, detect, and mitigate bias in AI-generated text.
• It provides a controlled Retrieval-Augmented Generation (RAG) pipeline, utilities for embedding and language detection, modules for different bias detectors, and integration with multiple LLM APIs (OpenAI GPT-4, Claude, Cohere, Mistral, DeepSeek).
File Structure Analysis
• Top-level directories and files:
• data/ – contains static datasets (e.g., retrieval_recency.csv)
• notebooks/ – Jupyter notebooks demonstrating usage and analysis
• src/
– queries/ – scripts to run specific data retrieval experiments
– unbiasai/
· biases/ – modules implementing bias detection logic
· config.py – environment and API key configuration
· utils.py – helper functions (embedding, LLM initialization, language detection)
• tests/ – unit tests for core utilities
• pyproject.toml & requirements.txt – dependency definitions
• .env or ENVFILE – environment variables (API keys, Supabase credentials)
Key Components and Their Relationships
a. Data Layer
– Static CSV datasets (data/)
– Supabase database for document retrieval (via create_client in utils/config)
b. Core Pipeline
– Retrieval module (src/queries and retrieval_bias code) fetches context from Supabase or CSV
– Embedding utility (get_embedding in utils.py) calls OpenAI embedding endpoint
– Language detection utility (detect_language in utils.py) identifies text language
– Bias detectors (modules in src/unbiasai/biases):
· retrieval_recency.py
· retrieval_language.py
· other bias types can be added here
– RAG orchestrator (can be in user code or notebooks) that:
1. Retrieves context
2. Generates or augments prompts via an LLM
3. Passes output to bias detection modules
c. External Services
– OpenAI API (embeddings, completions)
– Other LLM APIs (Claude, Cohere, Mistral, DeepSeek) via initialize_llm
– Supabase (vector DB or document store)
d. User Interface / Entry Points
– Python functions and classes in src/unbiasai/utils.py and src/unbiasai/biases
– Example scripts in src/queries/
– Jupyter notebooks demonstrating end-to-end analysis
Architectural Patterns and Design Principles
• Modular design: separate packages for utilities, biases, and query scripts
• Dependency inversion: utils.py abstracts LLM initialization and embeddings so bias modules are decoupled from specific LLM APIs
• Extensibility: new bias detectors can be added under src/unbiasai/biases; new retrieval strategies under src/queries
• Testability: core utilities have unit tests; business logic in pure Python modules
Elements
• Clear labels for each component (e.g., “Retrieval Module”, “Embedding Service”, “OpenAI API”, “Bias Detector: Retrieval Recency”)
– “Retrieval Module” → “Supabase (External DB)” and/or “CSV Data”
– “Embedding Service” → “OpenAI Embeddings API”
7. Surround everything with a boundary labeled “UnBiasAI Package (src/unbiasai).”