QingFei1/LongRAG

Implements a dual-perspective RAG architecture with separate Extractor and Filter components that decompose long-context understanding into global information retrieval and factual detail extraction. Built on LLaMA-Factory for supervised fine-tuning, it supports modular component composition—extractors and filters can be independently swapped across different LLM generators (ChatGLM3, Llama3, GPT-3.5, GLM-4). Evaluated on multi-hop QA datasets from LongBench with context lengths up to 32k tokens, achieving 52.56 F1 average with GLM-4.

120 stars. No commits in the last 6 months.