Hybrid Retrieval Architectures for Production RAG
About this session
Naive vector search demos beautifully and fails quietly in production. In regulated industries like banking and insurance, a wrong retrieval isn't just an inconvenience, it can be a compliance problem. This talk is based on lessons from building RAG systems for enterprise document intelligence and legacy modernization field mapping at large financial institutions. In 10 minutes I'll walk through the retrieval architecture that actually survived production: hybrid semantic and keyword search, cross-encoder reranking for precision, and an evaluation layer built on golden datasets, prompt versioning, and regression gates so that accuracy is auditable instead of anecdotal. You'll leave with a practical blueprint and a simple checklist for deciding when vector-only retrieval is enough and when it isn't.
Speaker
Key takeaways
- Why vector-only retrieval breaks on enterprise data and how hybrid semantic plus keyword search with reranking fixes it
- How to build an evaluation layer with golden datasets and regression gates so retrieval quality is measurable and auditable
- A practical checklist for choosing the right retrieval architecture based on data type, accuracy needs, and compliance requirements