Parses complex documents and images into a hybrid vector search and LightRAG knowledge graph with citations.
Combines vector search, LightRAG entity mapping, and vision LLMs to answer questions from visual documents.
NexusRAG is a retrieval-augmented generation engine that processes text, tables, and images from complex documents into a structured vector and graph database. It targets developers and AI engineering teams who need grounded answers backed by precise document citations. You can deploy it locally using Ollama and sentence-transformers or connect it to external providers like Gemini.
NexusRAG extracts layout structures, tables, and images using Docling or Marker, then generates visual descriptions via multimodal LLMs before chunking. It runs a three-way retrieval strategy across ChromaDB vector search, LightRAG entity traversal, and cross-encoder joint reranking. The React interface displays inline 4-character citations, document page previews, and force-directed knowledge graph network visualizers.
It suits software engineers, data scientists, and technical researchers building contextual search tools over complex PDFs, reports, and manuals. It is not intended for non-technical end users looking for a single-click desktop PDF reader. If you lack dedicated GPU hardware or API budget for vision models, document ingestion will be significantly slower.
Preserves heading hierarchies, page boundaries, and section relationships using configurable Docling or Marker parsing engines.
Appends vision LLM captions for images and structured markdown table summaries directly into chunk embeddings.
Executes ChromaDB vector over-fetching, LightRAG entity lookup, and cross-encoder reranking in parallel.
Generates four-character inline citation tags that link directly to page numbers and original heading paths.
Renders force-directed entity and relationship network graphs in the browser without requiring external graph database servers.
NexusRAG is free and open-source software under the MIT license. The repository author does not offer a paid managed cloud platform. Your operational costs depend entirely on self-hosted GPU infrastructure or commercial model API usage such as Gemini.
877 open remote AI-training roles, updated November 16, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Yes, you can configure NexusRAG to use local Ollama models and sentence-transformers, running parsing, embedding, and chat locally.
Depending on the active parser configuration, it accepts PDF, DOCX, PPTX, XLSX, HTML, EPUB, and TXT files.
Vision models generate text captions for extracted images during upload, which are embedded alongside the text on that page.
See Zekai first in Google