FlowForge
Multimodal RAG platform forked from RAGFlow with 70%+ rewrite, new UI, and VLM-powered retrieval.
Overview
FlowForge is a community-maintained fork of the upstream RAGFlow project that has diverged substantially: over 70% of the codebase has been rewritten, the entire web UI/UX has been replaced, and significant multimodal/VLM-powered capabilities have been added on top.
What’s different
- Brand-new front-end built around a window-management UX rather than a flat dashboard.
- Native support for vision-language models (VLMs) in the retrieval path — image chunks are first-class citizens alongside text.
- Re-engineered ingestion pipeline with pluggable chunkers and a stronger table extractor.
- Continued API compatibility with upstream RAGFlow where it makes sense.
Architecture highlights
The core stays Python (FastAPI) for the API layer and worker pool, with ElasticSearch for hybrid retrieval, MinIO for object storage, and Redis for the queue. The front-end is a TypeScript / React SPA that talks to a stable REST API. The VLM inference is dispatched to a configurable backend (open-weights served locally, or a managed provider via a thin adapter).
Initializing demo…