GenAI Core Framework: A Reusable Research Platform for Generative AI Experiments

GenAI Core Framework is a lightweight generative AI research framework designed to solve the problems of code duplication and irreproducible experiments in research. It provides standardized components for LLM applications, RAG systems, Agentic AI, and evaluation pipelines, supporting model providers like OpenAI and Ollama, helping researchers focus on innovation rather than infrastructure. The original author is Piyal Banik (Piyal-AI-Labs), and it was released on GitHub in June 2026.

Generative AI is developing rapidly, but researchers often face the problem of writing similar code from scratch for each project (e.g., LLM calls, RAG construction, evaluation metric implementation), leading to time wasted on duplicate code and poor experiment reproducibility—different projects have different implementation methods, making it difficult to compare results. GenAI Core Framework was created to address this core pain point.

Design Philosophy: Positioned as a research platform (not an end-to-end application framework), it follows these principles: Simple modularity (components can be used independently or combined), research-oriented (prioritizes experimental flexibility and reproducibility), reusable components (standardized interfaces for cross-project sharing), framework-agnostic (not tied to specific deep learning frameworks), and easy to extend (clear architecture for adding new features).

Core Components:

1. LLM Interface Layer: Unified access to providers like OpenAI (GPT series) and Ollama (local open-source models), allowing easy model switching without modifying upper-layer code;
2. Embedding Model Interface: Supports OpenAI Embeddings and Sentence Transformers, ensuring vector comparability;
3. Evaluation Metrics: Built-in standardized metrics such as Accuracy, BLEU, ROUGE, and BERTScore;
4. Experiment Tracking: Records hyperparameters, evaluation metrics, metadata, and experiment results to ensure reproducibility;
5. Configuration Management: YAML-based centralized management, with configuration as code to separate logic and settings;
6. Prompt Management: Supports version control, dynamic rendering (variable substitution/conditional logic), and A/B testing.

Project Architecture: Uses a layered structure with core directories including llms/ (LLM interfaces), embeddings/ (embedding model interfaces), evaluation/ (evaluation metrics), experiments/ (experiment tracking), configs/ (configuration management), and prompts/ (prompt templates), with clear responsibilities for easy understanding and extension.

Application Scenarios:

• LLM application development: Chatbots, content generation tools, code assistance tools;
• RAG system construction: Enterprise knowledge base Q&A, document retrieval and summarization, multi-source information integration;
• Agentic AI research: Autonomous agent systems, multi-agent collaboration, tool usage learning;
• Evaluation and benchmarking: Model performance comparison, new metric validation, systematic evaluation;
• Fine-tuning experiments: Domain adaptation, instruction fine-tuning, continuous learning;
• Multimodal AI research: Image-text understanding, cross-modal generation, multimodal retrieval.

Key differences between GenAI Core Framework and LangChain/LlamaIndex:

Feature	GenAI Core Framework	LangChain/LlamaIndex
Positioning	Research platform	Application framework
Complexity	Lightweight	Feature-rich but heavy
Flexibility	High (modular)	Medium (predefined workflows)
Learning curve	Gentle	Steeper
Production readiness	Requires customization	Relatively mature

GenAI Core Framework is more suitable for research-oriented projects, while mature application frameworks are better for quickly building production applications.

Future Plans: Soon to add features such as RAG-specific evaluation metrics, LLM-as-a-Judge evaluation methods, re-ranker support, retrieval tool sets, caching mechanisms, observability tools, and multimodal support.

Summary: The framework accurately addresses the pain points in generative AI research, with value in standardized interfaces (reducing switching costs), reproducibility (experiment tracking and configuration management), modularity (use as needed), and scalability (clear architecture). It is suitable for multi-experiment teams or researchers needing standardized processes, making work more organized and efficient, and serves as the foundation for all Piyal-AI-Labs projects.