Orchestrating Multimodal RAG With Agentic Workflows
summary
Built a multimodal RAG pipeline that combines OpenSearch MCP server, Quen2.5 VL (hosted on SageMaker AI) and AWS Strands as an orchestrator to let agents choose lexical/semantic/hybrid search, rerank, and call multimodal embedding tools. Uses Titan image embed v1 and VQA dataset. Claims: precision@1 100% on VQA, ~70% of queries return in <3s and median latency ~3.7s. Demonstrated ingestion, retrieval, and generation loops that converge on grounded answers.
AI-written summary · about
chapters (11)
- 0:00Multimodal agentic RAG intro
- 1:03Traditional RAG challenges
- 3:34Agentic orchestration benefits
- 7:07Multimodal RAG architecture
- 10:21Technologies powering the solution
- 13:51OpenSearch MCP server
- 15:59Dataset and embeddings
- 17:26Evaluation metrics and results
- 19:12Live demo
- 23:07Key takeaways
- 24:11Q&A
related talks
22mOrchestrating Multimodal RAG With Agentic Workflows With OpenSearch
34mImplementing an open-source RAG with OpenSearch
01h 28mFrom RAG to Riches Building Enterprise Search Systems with LLMs
01h 30mOptimizing Retrieval for Agentic Systems
25mSet Up an Open Source RAG in 30 Minutes
31mBuild an Agentic RAG system using Langchain, Ollama and Milvus
59mRAG Against the Machine
21mBeyond RAG: Building Robust Knowledge Frameworks for Complex Multi-document Retrieval for Enterprise