Documents are images, not text, to this agent stack

What it does

LAYRA is a document-centric automation platform that treats PDFs, slides, and spreadsheets as visual artifacts rather than stripping them to plain text. It ingests over 100 formats, converts pages to image embeddings using ColQwen 2.5 or Jina-Embeddings-v4, and stores vectors in Milvus for retrieval. A Next.js frontend and FastAPI backend expose a drag-and-drop workflow builder where nodes can run sandboxed Python, call LLMs, loop, branch, and pause for human approval.

The interesting bit

Traditional RAG shreds documents into token soup and loses tables, charts, and layout; LAYRA embeds the intact page image so the model “sees” the document like a human reader. That visual retrieval feeds directly into a workflow engine supporting nested loops, breakpoint debugging, and MCP tool calls—turning document pipelines into a visual programming environment.

Key highlights

• Visual RAG via ColQwen 2.5 (local, needs a >16 GB GPU and ~15 GB of weights) or Jina-Embeddings-v4 API (cloud, no GPU required)
• Workflow engine with human-in-the-loop gates, chat memory, and sandboxed Python execution
• Enterprise microservice stack: FastAPI, Redis, MySQL, MongoDB, Kafka, MinIO, and Milvus
• Supports 100+ formats including PDF, DOCX, XLSX, and PPTX
• Real-time SSE streaming of workflow execution with node-level debugging and reusable components

Caveats

• Local embedding mode demands serious hardware: a GPU with more than 16 GB of VRAM and roughly 15 GB of downloaded model weights
• The visible quickstart assumes containerized deployment; bare-metal setup is not discussed

Verdict

Worth a look if your team is drowning in complex documents—forms, diagrams, and scanned layouts—that traditional text-RAG mangles. Skip it if you just need a lightweight text-only Q&A bot; the infrastructure footprint is commensurately heavy.