Claude Code + RAG-Anything = LIMITLESS

하이라이트

• RAG-Anything enables LightRAG to process non-text documents including scanned PDFs, charts, graphs, and LaTeX equations.

• The system uses MinerU, an open-source local document parser, to decompose documents into headers, text blocks, and screenshots of visual elements.

• Processing costs are reduced by using local models like PaddleOCR for text extraction while reserving high-tier LLMs like GPT-4o-mini for entity and relationship extraction.

• Non-text documents follow a dual-path pipeline that creates separate vector databases and knowledge graphs for text and images before merging them into a single searchable index.

• Integration with Claude Code allows users to ingest complex documents via a single terminal command using a custom 'rag-anything' skill.

• A Novatech SAS revenue analysis test demonstrated the system's ability to accurately extract and query monthly revenue trends from a visual bar chart.

타임라인

1. 00:00 - 01:38

   Limitations of Text-Centric RAG Systems

   Most existing RAG solutions are restricted to digital text formats, making them ineffective for real-world business documents that rely on visual data representation. LightRAG specifically struggles with these formats despite its advanced knowledge graph capabilities. RAG-Anything addresses this by plugging directly into the existing LightRAG stack to expand its ingestion capabilities to include any document format.

   • Standard Retrieval-Augmented Generation (RAG) systems fail when encountering non-textual data like charts and graphs.
   • Scanned PDF documents are often unreadable to text-based RAG architectures like LightRAG.
   • RAG-Anything acts as a specialized wrapper to bridge the gap between visual documents and text-based knowledge graphs.
2. 01:39 - 04:36

   Local Document Parsing with MinerU

   The big-picture goal of this system is to look at documents that are not natively text and break them down into their fundamental parts. MinerU is an open-source project that utilizes specialized miniature models to categorize every element on a page. This local step ensures that subsequent processing is handled by the most efficient model for each specific data type.

   • MinerU runs locally on the user's computer to identify and box different document components.
   • The parser distinguishes between headers, standard text, charts, bar graphs, and mathematical equations.
   • Understanding document architecture is a prerequisite for creating custom AI skills and advanced developer projects.
3. 04:37 - 09:51

   The Dual-Path Ingestion Pipeline

   The system splits data into two buckets: text and images. The text bucket is sent to an LLM like GPT-4o-mini to extract entities and relationships for a knowledge graph and generate embeddings for a vector database. Simultaneously, the image bucket is sent to a vision model to perform the same extraction on screenshots. This parallel processing ensures that information hidden in charts is treated with the same structural importance as written prose.

   • Textual components are processed via PaddleOCR to convert scanned pixels into readable strings.
   • Visual elements that cannot be converted to text are captured as screenshots for vision-capable LLMs.
   • A scalpel-like approach to data extraction prevents the high costs and slow speeds associated with full-page OCR via expensive APIs.
4. 09:52 - 13:07

   Merging Knowledge Graphs and Vector Databases

   After the LLM processes both the text and image buckets, it generates two sets of embeddings and relationship maps. These are merged based on shared entities to create a comprehensive representation of the document. For users already running a LightRAG server, this new RAG-Anything data is further merged into the existing global knowledge graph, making the entire ingestion process invisible during the query phase.

   • Four distinct data objects are created from a single non-text document: two vector databases and two knowledge graphs.
   • Merging these objects results in a single, unified searchable index where text and visual data overlay.
   • The final merged database allows for natural language queries that draw from both textual descriptions and visual chart data.
5. 13:08 - 16:13

   Installation and Claude Code Integration

   Setting up the environment involves updating storage paths and fixing common 'embedding double wrap' bugs found in the original GitHub repository. Because RAG-Anything operates as a wrapper, it requires a separate pathway for ingestion compared to standard text files. By converting the ingestion script into a Claude Code skill, users can upload complex documents simply by pointing the AI to a specific file or folder in the terminal.

   • A one-shot prompt in Claude Code automates the installation of MinerU and its heavy dependencies.
   • The system defaults to running on the CPU but can be transitioned to the GPU by updating PyTorch versions.
   • Standard LightRAG UIs do not support non-text uploads, requiring a Python script or custom Claude Code skill.
6. 16:14 - 19:22

   Practical Application and Revenue Analysis Test

   The system was tested using a Novatech SAS revenue PDF containing visual charts that a standard RAG would ignore. By using a plain language query, the system accessed the merged knowledge graph to provide a detailed monthly breakdown of revenue trends from January to September 2025. This demonstrates that once the initial architectural hurdle of image parsing is cleared, the user experience remains identical to standard text-based AI interactions.

   • A query regarding a fake company's SAS revenue successfully extracted data points directly from a bar chart.
   • The system returned specific monthly figures, such as $4.6M for January and $5.4M for March, which were only present in image form.
   • Claude Code automatically manages retrieval parameters and API structures for local RAG instances.