G4-RAG: A Retrieval-Augmented Generation System Enhanced with Adaptive Chunking and Agentic Workflow

G4-RAG is an improved retrieval-augmented generation system that proposes optimization solutions for the pain points of traditional RAG architectures. Core improvements include: adopting an adaptive chunking strategy to solve the problem of document semantic integrity; using FAISS vector retrieval + cosine similarity re-ranking to balance efficiency and retrieval quality; extending the Agentic workflow via Pydantic AI to support multi-step reasoning and tool calls; using ROUGE and BERTScore for system evaluation. The project aims to build a more robust and efficient RAG system, providing reliable solutions for practical applications.

Retrieval-Augmented Generation (RAG) is one of the mainstream architectures for LLM applications, which alleviates the issues of knowledge cutoff and hallucinations by integrating external knowledge retrieval. However, traditional RAG faces challenges such as document chunking granularity selection, relevance ranking of retrieval results, and maintenance of multi-turn dialogue context. G4-RAG proposes improvement solutions for these pain points to enhance system performance.

Traditional fixed-length chunking easily breaks semantic integrity (e.g., splitting paragraphs/code blocks). G4-RAG adopts adaptive chunking: dynamically adjusts splitting based on document structure (natural boundaries like paragraphs, chapters, code blocks); identifies title hierarchies for structured documents; analyzes information density to balance chunk size, achieving a balance between integrity and retrieval accuracy.

G4-RAG adopts a two-stage retrieval strategy: the first stage uses FAISS to quickly recall candidate chunks (efficient approximate nearest neighbor search); the second stage uses cosine similarity re-ranking to finely evaluate semantic relevance. This strategy not only ensures response speed but also improves the quality of retrieval results, making up for the boundary limitations of FAISS approximate search.

Traditional RAG is mostly a single-turn process. G4-RAG implements an Agentic workflow via Pydantic AI: supports multi-step reasoning (determines single/multi-turn retrieval based on query complexity); can call external tools (web search, database query, etc.) to expand capabilities; Pydantic AI provides type-safe Agent definitions, reducing development and maintenance costs.

G4-RAG uses ROUGE (n-gram overlap) and BERTScore (semantic similarity) for evaluation: ROUGE reflects surface faithfulness, while BERTScore captures semantic equivalence. Evaluation results verify the effectiveness of adaptive chunking, two-stage retrieval, and Agentic extension—especially in complex queries, the Agentic workflow significantly improves the completeness and accuracy of answers.

G4-RAG is suitable for enterprise knowledge base Q&A (processing long documents + complex business problems), academic research assistance (literature location + comprehensive information generation), and customer service (fast response + multi-knowledge base query). The open-source implementation of the project provides reusable components to help developers enhance the capabilities of existing RAG systems.

The improvement directions of G4-RAG (adaptive chunking, two-stage retrieval, Agentic extension) are mutually synergistic: chunking lays the foundation for retrieval, precise retrieval supports Agentic reasoning, and Agentic capabilities guide intelligent retrieval. The project provides a reference for the evolution of RAG architectures and demonstrates the value of engineering optimization and architectural innovation for knowledge-enhanced systems.