Transforming Banking with Advanced AI: GenAI, RAG, and Multi-Agent Systems
summary
Belfius builds a production RAG pipeline that ingests Confluence content, cleans messy JSON/HTML, chunks documents, creates multilingual embeddings and stores them in an ElasticSearch vector DB, then reranks results and sends contextual chunks to an LLM. Demoed use: fraud FAQ search returns one consolidated, hyperlinked answer instead of surfacing multiple pages. Measures tracked: precision/recall for retrieval and relevancy/faithfulness for generation. Main failures are poor source quality and token limits; guardrails, retraining data standards and reranking fix most issues.
AI-written summary · about
chapters (16)
- 0:00Introduction to AI at Belfius
- 1:28History and evolution of AI
- 5:24Predictive vs generative AI
- 7:13Understanding RAG principles
- 10:41Data quality for RAG
- 17:26Evaluating RAG performance
- 19:40Key takeaways for RAG
- 20:51RAG architecture components
- 23:24Chunking and reranking
- 30:54RAG guardrails and security
- 33:04RAG demo and multilingualism
- 38:54Advanced prompt engineering
- 41:37Future trends in RAG
- 45:19Hypothetical embeddings
- 46:38Multi-agent systems
- 50:31Closing remarks
related talks
28mImproving responses in generative AI applications
27mTransforming Financial Services with Generative AI
21mHow RAG is Redefining AI Capabilities
15mRAG Beyond Chatbots
01h 16mForget RAG Pipelines—Build Production Ready Agents in 15 Mins
40mBuilding Intelligent Search Systems With Agentic RAG and OpenSearch
36mGENAI-Powered chatbot to increase team efficiency
21mSmart Search with RAG: Elasticsearch Meets Language Models