Agentic RAG Frameworks

WIP nlp-llms agents graph-rag lightrag llm llm-agents nlp rag retrieval 1 min read

Agentic RAG, self-reflection, iterative retrieval, and graph-based frameworks (LightRAG, RAG-Anything)

Overview

Unlike standard linear RAG pipelines, RAG Agents actively decide when to search, how to formulate queries, and when to reflect on retrieved context. This includes hybridizing traditional vector search with Knowledge Graphs (KG).

Advanced RAG Frameworks

LightRAG

LightRAG is a lightweight knowledge-graph RAG framework and an efficient alternative to Microsoft GraphRAG.

Dual-Layer Architecture: Bridges the gap between traditional vector-based RAG and graph-based RAG by managing both Knowledge Graphs and vector embeddings.
Multi-hop Subgraphs: Extracts global information from constructed graph structures, drastically enhancing the model’s ability to handle complex queries that span multiple document chunks.

RAG-Anything

RAG-Anything acts as an all-in-one RAG framework, highly extensible for various data sources, designed by the HKUDS lab.

Self-Reflection (CRITIC / Self-RAG)

TODO: Add details.

Query Routing

TODO: Add details.