Loading repository data…
Loading repository data…
liweiphys / repository
LAYRA—an enterprise-ready, out-of-the-box solution—unlocks next-generation intelligent systems powered by visual RAG and limitless visual multi-step agent workflow orchestration.
A transparent discovery signal based on current public GitHub metadata.
This score does not audit code, security, maintainers, documentation quality, or suitability. Verify the repository and its current documentation before adoption.
🚀 New Jina-Embeddings-v4 API support eliminates local GPU requirements
LAYRA is the world’s first “visual-native” AI automation engine. It sees documents like a human, preserves layout and graphical elements, and executes arbitrarily complex workflows with full Python control. From vision-driven Retrieval-Augmented Generation (RAG) to multi-step agent workflow orchestration, LAYRA empowers you to build next-generation intelligent systems—no limits, no compromises.
Built for Enterprise-Grade deployment, LAYRA features:
Explore LAYRA's powerful interface and capabilities through these visuals:
Homepage - Your Gateway to LAYRA
Knowledge Base - Centralized Document Hub
Interactive Dialogue - Layout-Preserving Answers
Workflow Builder - Drag-and-Drop Agent Creation
Workflow Builder - MCP Example
Before starting, ensure your system meets these requirements:
# Clone the repository
git clone https://github.com/liweiphys/layra.git
cd layra
# Edit configuration file (modify server IP/parameters as needed)
vim .env
# Key configuration options include:
# - SERVER_IP (server IP)
# - MODEL_BASE_URL (model download source)
For Jina (cloud API) Embeddings v4 users:
vim .env
EMBEDDING_IMAGE_DPI=100 # DPI for document-to-image conversion. Recommended: 100 - 200 (12.5k - 50k tokens/img)
EMBEDDING_MODEL=jina_embeddings_v4
JINA_API_KEY=your_jina_api_key
JINA_EMBEDDINGS_V4_URL=https://api.jina.ai/v1/embeddings
Option A: Local ColQwen deployment (recommended for GPUs with >16GB VRAM)
# Initial startup will download ~15GB model weights (be patient)
docker compose up -d --build
# Monitor logs in real-time (replace <container_name> with actual name)
docker compose logs -f <container_name>
Option B: Jina-embeddings-v4 API service (for limited/no GPU resources)
# Initial startup will not download any model weights (fast!)
docker compose -f docker-compose-no-local-embedding.yml up -d --build
# Monitor logs in real-time (replace <container_name> with actual name)
docker compose logs -f <container_name>
Note: If you encounter issues with
docker compose, try usingdocker-compose(with the dash) instead. Also, ensure that you're using Docker Compose v2, as older versions may not support all features. You can check your version withdocker compose versionordocker-compose version.
Your deployment is complete! Start creating with Layra now. 🚀✨
For detailed options, see the Deployment section.
📘 Essential Learning: We strongly recommend spending just 60 minutes with the tutorial before starting with LAYRA - this small investment will help you master its full potential and unlock advanced capabilities.
For step-by-step instructions and visual guides, visit our tutorial on GitHub Pages:
Tutorial Guide
While LAYRA's Visual RAG Engine revolutionizes document understanding, its true power lies in the Agent Workflow Engine - a visual-native platform for building complex AI agents that see, reason, and act. Unlike traditional RAG/Workflow systems limited to retrieval, LAYRA enables full-stack automation through:
🔄 Cyclic & Nested Structures
Build recursive workflows with loop nesting, conditional branching, and custom Python logic - no structural limitations.
🐞 Node-Level Debugging
Inspect variables, pause/resume execution, and modify state mid-workflow with visual breakpoint debugging.
👤 Human-in-the-Loop Integration
Inject user approvals at critical nodes for collaborative AI-human decision making.
🧠 Chat Memory & MCP Integration
Maintain context across nodes with chat memory and access live information via Model Context Protocol (MCP).
🐍 Full Python Execution
Run arbitrary Python code with pip installs, HTTP requests, and custom libraries in sandboxed environments.
🎭 Multimodal I/O Orchestration
Process and generate hybrid text/image outputs across workflow stages.
Traditional RAG systems fail because they:
LAYRA changes this with pure visual embeddings:
🔍 It sees each page as a whole - just like a human reader - preserving:
- ✅ Layout structure (headers, lists, sections)
- ✅ Tabular integrity (rows, columns, merged cells)
- ✅ Embedded visuals (plots, graphs, stamps, handwriting)
- ✅ Multi-modal consistency between layout and content
Together, these engines form the first complete visual-native agent platform - where AI doesn't just retrieve information, but executes complex vision-driven workflows end-to-end.
Code Without Limits, Build Without Boundaries Our Agent Workflow Engine thinks in LLM, sees in visuals, and builds your logic in Python — no limits, just intelligence.
🔄 Unlimited Workflow Creation
Design complex custom workflows without structural constraints. Handle unique business logic, branching, loops, and conditions through an intuitive interface.
⚡ Real-Time Streaming Execution (SSE)
Observe execution results streamed live – eliminate waiting times entirely.
👥 Human-in-the-Loop Integration
Integrate user input at critical decision points to review, adjust, or direct model reasoning. Enables collaborative AI workflows with dynamic human oversight.
👁️ Visual-First Multimodal RAG
Features LAYRA’s proprietary pure visual embedding system, delivering lossless document understanding across 100+ formats (PDF, DOCX, XLSX, PPTX, etc.). The AI actively "sees" your content.
🧠 Chat Memory & MCP Integration
🐍 Full-Stack Python Control
🎨 Flexible Multimodal I/O
Process and generate text, images, or hybrid outputs – ideal for cross-modal applications.
🔧 Advanced Development Suite
🧩 Intelligent Data Utilities
Forget tokenization. Forget layout loss.
With pure visual embeddings, LAYRA understands documents like a human — page by page, structure and all.
LAYRA uses next-generation Retrieval-Augmented Generation (RAG) technology powered by pure visual embeddings. It treats documents not as sequences of tokens but as visually structured artifacts — preserving layout, semantics, and graphical elements like tables, figures, and charts.
(2025.8.4) ✨ Expanded Embedding Model Support:
colqwen (Local GPU - high performance)jina-embeddings-v4 (Cloud API - zero GPU requirements)(2025.6.2) Workflow Engine Now Available: