Awesome-LLM-RAG: A Panoramic Paper Repository for Retrieval-Augmented Generation (RAG) Technology

Awesome-LLM-RAG is a carefully curated list of RAG technology papers, covering the complete research trajectory from basic architecture to cutting-edge optimizations. It builds a knowledge graph using a multi-dimensional classification system, providing systematic academic navigation for LLM application developers and researchers, and addressing the problem of information overload in the RAG field.

Retrieval-Augmented Generation (RAG) has become the de facto standard for LLM application development, effectively alleviating the knowledge timeliness and hallucination issues of purely parametric models. However, the RAG field evolves rapidly (from its 2020 foundation to cutting-edge developments like adaptive retrieval and GraphRAG), leading to information overload for researchers—hence the birth of the Awesome-LLM-RAG project.

The project features a multi-dimensional classification system, divided into more than ten categories based on research topics and application scenarios: At the basic architecture level, it includes foundational RAG papers (e.g., the 2020 NeurIPS paper Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks); At the optimization technology level, it is subdivided into directions like instruction fine-tuning, in-context learning, embedding models, and search strategies, reflecting RAG's evolution path from simple architectures to complex systems.

The project focuses on four core directions: 1. Adaptive retrieval (models determine whether external knowledge is needed, e.g., Self-RAG, FLARE); 2. Integration of retrieval and long context (exploring the necessity and integration methods of RAG in long-context models); 3. Instruction fine-tuning optimization (e.g., RA-DIT improves models' ability to utilize retrieval results); 4. GraphRAG (introducing knowledge graphs to address multi-hop reasoning limitations).

In addition to academic papers, the project includes RAG-themed sessions from top conferences (e.g., 2023 ACL tutorial, SIGIR workshop) and practical guides (such as Build an Advanced RAG Application and Enterprise RAG), covering the knowledge system from theory to enterprise-level deployment.

The evolution trajectory of RAG technology is clear: 2020-2022 focused on basic architecture; 2022-2023 focused on retrieval strategy optimization and instruction fine-tuning; 2024-2025 shifted to cutting-edge areas like adaptive mechanisms and multi-modal expansion. This evolution reflects the transformation process from academic research to industrial applications.

Researchers can quickly locate literature and clarify their research positioning; algorithm engineers can obtain practical details and reference evaluation methods; technical decision-makers can assess technology maturity and trends, helping with selection and investment strategies.

The project uses the MIT license, encouraging community contributions (via PR submissions) to ensure timeliness; it provides a convenient navigation structure to enhance the user experience. In the future, RAG will expand to multi-modal knowledge fusion and dynamic knowledge updates, and this repository serves as a valuable starting point for deepening one's understanding of the field.