Developers can build reliable AI systems by converting complex documents into usable context
RAGFlow organizes information scattered across PDFs, tables, and images so that AI can understand and answer accurately. It goes beyond simple retrieval by automating the process where an agent analyzes questions, finds necessary information, and verifies it. This allows developers to build trustworthy AI systems quickly without writing complex data preprocessing code.
Agentic search in version 1.0
The release of RAGFlow 1.0.0-rc1 clarified the project's direction. This update introduced 'agentic RAG,' enabling the model to analyze questions, decompose them as needed, retrieve knowledge, and verify evidence in a multi-step process. It also added 'knowledge compilation,' which generates structured knowledge outputs like wikis, graphs, trees, and mind maps at the document and dataset levels. These changes show that RAGFlow has evolved from a simple search tool into context infrastructure for LLMs.
Document understanding and knowledge structuring
The core of RAGFlow is extracting knowledge from complex, unstructured data based on deep document understanding. It supports various data sources including Word, PowerPoint, Excel, PDF, scans, images, and web pages, performing intelligent and explainable segmentation through template-based chunking. The 'knowledge compilation' feature allows converting document content into reusable formats such as wikis, graphs, trees, PageIndex, mind maps, timelines, and skills. This creates structured knowledge outputs rather than simple text fragments, contributing to higher AI answer quality.
Verifiable answers and reduced hallucination
The system is designed to track the origin of AI-generated answers. It visualizes the text chunking process to allow human intervention and provides key references and traceable citations to support evidence-based responses. This reduces hallucinations and raises the trustworthiness of AI answers for users. Additionally, it offers four thinking modes—Low, Medium, High, and Ultra—to adjust the depth of search and reasoning according to question complexity.
Local deployment and system requirements
RAGFlow supports two deployment methods: via Docker and directly from source. Docker deployment is suitable for quick evaluation, integration testing, and production, and does not require a Go compiler. The recommended starting configuration is 4 CPU cores, 16GB RAM, and 50GB disk space, though actual requirements vary based on the document engine, data volume, parsing tasks, and concurrency. Docker version 24.0.0 or higher and Docker Compose v2.26.1 or higher are required. Running from source requires the Go version specified in go.mod (currently Go 1.27), Clang 20, LLD 20, CMake 4.0 or higher, and PCRE2 development files. macOS is not currently supported for Docker deployment.
Considerations before adoption
In the 1.0 release stage, note that DeepDoc uses CPU inference for layout analysis, OCR, and table recognition. This indicates performance in environments without GPU acceleration, so processing time should be considered for large-scale document handling. When using Elasticsearch, the vm.max_map_count on the Docker host must be set to 262144 or higher; since this setting resets upon system reboot, it is recommended to add it permanently to /etc/sysctl.conf. You must configure the names, service addresses, and API keys for LLM, embedding, and reranker models on the model provider page before the service can be used. As an open-source project, it is suitable for teams capable of solving problems through community contributions and documentation.
