Most RAG stacks quietly throw away half of every document. The moment a PDF holds a chart, a results table, or a LaTeX equation, a text-only retriever either skips it or flattens it into a meaningless string — yet that is often exactly where the answer lives. The bet here is that multimodal content deserves first-class treatment inside the knowledge graph, not a bolt-on caption.
What Sets It Apart
- Built directly on LightRAG, so it inherits graph-based retrieval but adds parallel channels for images, tables, and formulas — a query can traverse from a paragraph to the chart that actually supports it.
- A five-stage pipeline (parse → analyze → build graph → retrieve → rank) where parsing is pluggable: MinerU, Docling, or PaddleOCR. You pick the parser that copes with your document quirks instead of being locked into one.
- Retrieval fuses vector similarity with graph traversal, then re-ranks by modality, so a question about a number returns the relevant table cell rather than the prose around it.
Who It's For
Great fit if you index research papers, technical manuals, or enterprise reports where tables and figures carry the real signal, and you want one framework instead of a dozen glued-together tools. Look elsewhere if your corpus is plain prose — the multimodal graph and parser overhead buy you little — or if you need a turnkey hosted service rather than a Python framework you assemble and run yourself.
