Key takeaways
- Smart Document Processing: MariaDB AI RAG 1.1 now understands complex document layouts like tables and columns, ensuring your AI gets high-quality, structured information instead of messy text.
- Effortless Cloud Syncing: You can now connect directly to S3 storage to automatically sync only new or updated files, saving you significant time and computing costs.
- Superior Answer Accuracy: By using new "reranking" tools from Cohere and Flashrank, the system picks the most relevant data for the AI, which drastically cuts down on incorrect or "hallucinated" answers.
- One-Click Deployment: The entire RAG pipeline is now containerized with Docker, allowing you to launch the full environment—from document parsing to the vector database—with a single command.
Moving a RAG (Retrieval-Augmented Generation) application from a local prototype to a production-grade system requires solving for data messy ingestion, retrieval precision, and deployment complexity. MariaDB AI RAG 1.1 addresses these bottlenecks with improvements to the ingestion pipeline, more granular retrieval controls, and a containerized deployment model.
MariaDB AI RAG is currently in beta. You can evaluate the full stack with a trial license.
Layout-Aware Text Extraction
Plain text extraction often fails on complex documents, losing the context of tables, multi-column layouts, and headers. This leads to poor chunking and lower-quality embeddings.
MariaDB AI RAG 1.1 natively integrates Docling and LlamaParse to provide layout-aware extraction. By preserving the layout and structure of documents, the system generates cleaner chunks, directly improving the quality of the retrieved context and the resulting LLM response.
Bulk Document Ingestion from S3
Managing large document libraries requires efficient synchronization. You can now ingest document collections directly from S3-compatible object storage.
- Incremental Sync: The system tracks changes and only processes new or modified documents.
- Efficiency: This avoids the cost and compute overhead of re-indexing your entire knowledge base every time a file is added.
Precision Retrieval with Cross-Encoder Reranking
Vector search is fast but may not always rank the most relevant results at the top. To improve retrieval precision, MariaDB AI RAG 1.1 introduces a reranking step:
- Cohere Integration: Use Cohere’s managed reranker for high-quality, enterprise-grade ranking.
- Built-in Flashrank: For teams requiring lower latency or local processing, Flashrank is included as a lightweight, built-in option.
Reranking helps to refine the top results to ensure the LLM receives the most relevant context, significantly reducing hallucinations.
Docker Deployment
To simplify the setup of the RAG pipeline, version 1.1 is available as a containerized stack using Docker. With a single Docker command, you can deploy all RAG pipeline components with a single, repeatable step:
- The RAG pipeline components.
- The document parsing engine (Docling).
- The MCP (Model Context Protocol) Server.
- An optional MariaDB vector store for a complete end-to-end evaluation environment.
Getting Started
MariaDB AI RAG 1.1 delivers cleaner ingestion, better retrieval precision, and more transparent outputs through citations. These capabilities allow teams to validate results and iterate faster. To test it hands-on, MariaDB AI RAG is available for evaluation through a trial license.
