Rio Tinto unlocks potential of generative AI
summary
Rio Tinto builds a hybrid RAG pipeline on Amazon Bedrock using Titan embeddings, Neo4j-style knowledge graphs, OpenSearch-based vector indices, Lambdas for ingestion, and ECS for hybrid retrieval and generation. The system fuses graph traversal with semantic search, pulls 11–17 technical documents per query vs 2–3, and reports a 53% reduction in standard deviation on context and entity recall over 100 industry questions judged by a large LLM. The assistant runs deterministically within a fixed search radius and materially improves technical depth and recall.
AI-written summary · about
chapters (9)
- 0:00Introduction to hybrid RAG
- 3:55Limitations of traditional RAG
- 7:51Graph RAG and hybrid approach
- 9:57Rio Tinto's hybrid RAG implementation
- 13:32System architecture details
- 20:12Hybrid RAG demo and results
- 23:02Technical decisions behind design
- 25:11Enterprise search patterns with hybrid RAG
- 25:33Conclusion and recommendations
related talks
40mBuilding Intelligent Search Systems With Agentic RAG and OpenSearch
38mGenerative AI - Architectures and applications in depth
29mCustomizing genAI applications for your business
20mBuilding a secure end-to-end generative AI application in the cloud
23mNo RAGrets
22mGenerative AI application stack and providing long term memory to LLMs
24mBuilding GenAI applications with Amazon Bedrock
20mExplore what’s possible with new AWS generative AI services