v1.0.2

What's Changed

• Eml loader update by @EnjoyBacon7 in #40
• Properly handle special characters in partition name and file id by @slivering in #26
• added actor management endpoint by @htagourti in #45
• /complement answer improvement by @Ahmath-Gadji in #47
• Gh nightly workflow by @EnjoyBacon7 in #46
• Benchmark updates by @antncy in #52
• Fix chunks by @Ahmath-Gadji in #54
• Set explicit vllm version by @dodekapod in #62
• Doc cleaning by @Ahmath-Gadji in #73

New Contributors

• @slivering made their first contribution in #26

Full Changelog: v1.0.1...v1.0.2

v1.0.1

🚀 New Features

📐 Improved Chunking Logic

The Chunker class has been refactored to treat tables and images as standalone chunks instead of splitting them. This improves the quality of both retrieval results and LLM-generated responses, especially for structured or multimedia-rich documents.

⚙️ Ansible Playbooks for Remote Deployment

Introduced deployment script and ansible playbooks to support automated remote deployment of OpenRAG. Easily provision and deploy OpenRAG across remote nodes or cloud instances with minimal manual setup.

Full Changelog: v1.0.0...v1.0.1

v1.0.0

🚀 New Features

🔁 Ray Serve Integration

You can now use Ray Serve as an alternative to Uvicorn for serving HTTP requests.

📊 Milvus Integration Overhaul

Milvus is now managed as a Ray actor, enabling higher concurrency and better resource handling compared to the prior proxy-based setup.
Integration now uses pymilvus directly, removing the dependency on langchain_milvus. This simplifies maintenance and unlocks access to the latest Milvus features.

🛠 Enhancements & Fixes

• Improved Ray Actor Stability
  • Fixed crashes and out-of-memory (OOM) issues.
  • Actors now initialize earlier to reduce creation latency.
• Queue & Serialization Improvements
  • Fixed queue freezing issues.
  • Added timeout and retry logic to serialization tasks to better handle intermittent errors.
  • Tasks are now queued before dispatching to Ray workers, offering better queue visibility and load distribution.
• GlusterFS Support for Shared Storage
  Switched from NFS to GlusterFS for shared storage across Ray nodes.
• Use of PostgreSQL instead of SQLite
• Streamlined Cluster Operations
  • Auto-start Ray workers—no need for manual SSH into worker nodes.
  • Auto-restart of semaphore actors improves stability of coordination mechanisms across the system.
• Fixed sync issues with UV when running outside Docker

Full Changelog: v0.1.0...v1.0.0

v0.1.0

RAGondin v0.1.0 Release Notes

RAGondin is a lightweight, modular, and fully open-source Retrieval-Augmented Generation (RAG) framework designed for advanced RAG experimentation with a focus on sovereignty, scalability, and extensibility.

Key Features

Advanced Document Processing

• Supports a wide range of document formats including text, PDF, Office files, audio/video (via Whisper), and images with VLM-generated captions.
• Extracts and processes text, tables, images, and charts to provide deeper insights and richer contextual understanding.
• Flexible chunking strategies: recursive, semantic, and markdown chunkers with optimized default settings.

Scalable & Modular Architecture

• Designed for scalable deployment in distributed environments using Ray clusters for high-throughput parallel processing.
• Modular and flexible architecture enables easy customization and extension to meet diverse use cases.

Optimized Data Management & Security

• Vector indexing powered by Milvus with state-of-the-art multilingual embeddings (default: Qwen3-Embedding).
• Hybrid semantic and keyword search combining BM25 with Reciprocal Rank Fusion for superior retrieval.
• Data partitioning and secure access control designed for large organizations and sensitive environments.
• Token-based authentication for secure API access.

Seamless Integration & User Interfaces

• Provides RESTful APIs for indexing, searching, and RAG pipelines.
• OpenAI-compatible API and chat interface for easy adoption and integration with existing OpenAI-based workflows.
• Chainlit UI: Interactive chat interface with optional authentication and chat history persistence.
• IndexerUI: Web interface for intuitive document ingestion, indexing, and management.

Deployment

• Containerized deployment with Docker Compose, supporting both GPU and CPU environments for maximum flexibility.

Getting Started

• Requires Python 3.12+, Docker, and NVIDIA Container Toolkit for GPU acceleration.
• Configurable via Hydra and environment variables for embedding, retrieval, reranking, and transcription models.
• Ready-to-use APIs and interfaces allow quick setup for both development and production environments.