LLM-as-a-Judge and ReAct Agent: Chinese Culinary Arts RAG

🥗 This project dives into Retrieval-Augmented Generation (RAG) with a LLM-as-a-Judge evaluation and then empowered by ReAct agent, focused on Chinese cuisine, cooking, culture, and ingredients.

🥗 Built with LangChain, it integrates LLMs like Google’s Gemini, Kimi K2, and DeepSeek V3, paired with Qdrant Cloud for vector storage and Gemini/IBM Granite embeddings. The LLM-as-a-Judge system scores RAG outputs for quality, while the LLM agent delivers expert insights enhanced by reasoning–action–observation and tools.

🥗 For enthusiasts, housewives, or businesses exploring Chinese cuisine, this project offers a robust solution. It acts as Chinese food and culture expert and provides an intelligent agent that answers queries about Chinese recipes, cooking techniques, cultural traditions, and ingredients, making domain-specific knowledge accessible and engaging.

🥗 The pipeline builds multiple RAG setups with varied LLMs and embeddings. The LLM-as-a-Judge framework, inspired by pairwise comparison like ChatArena, evaluates outputs using rubrics (relevance, consistency, context fit, suitability). The ReAct agent runs in a reasoning-action-observation loop, retrieving data from web search, Wikipedia, and RAG knowledge base.

🥗 The LLM-as-a-Judge uses a scoring system for fair evaluation:
- Clear Ranking: Best answer earns 1 point, second-best 0.5 point, third 0 point.
- All Equally Good: If all answers are of similar high quality, each gets 0.33 point.
- Two Tied for Best: If two answers outperform the third, each gets 0.5 point; third gets 0.
* Explanations justify each score, ensuring transparency.

Tech Stack:
- RAG Pipeline: Python, LangChain, Qdrant, Google AI Studio API, OpenRouter API, HuggingFace
- Embedding Models: Gemini Embedding, IBM Granite Embedding
- Vector Database: Qdrant Cloud
- Techniques: LLM-as-a-Judge, Pairwise Comparison, Agentic RAG, Document Chunking