NexusRAG

Why we included this project

For teams building RAG over documents that are messy, dense, or full of tables, single-vector search rarely cuts it. NexusRAG attacks the problem three ways: vector similarity, knowledge-graph entity lookup through LightRAG, and cross-encoder reranking, run in parallel and combined into one answer. It also handles document parsing more seriously than most minimal demos, with Docling preserving headings and layout and a vision model captioning images and tables so they become searchable instead of being discarded. The streaming chat answers with inline citations, so you can trace every claim back to a source. Because it can run on local Ollama models or fully offline sentence-transformers, evaluating it for a document-heavy knowledge base does not require a cloud API.