Large Language Models (LLMs) are powerful for reasoning and natural language generation, but their effectiveness depends on access to fresh, relevant data. Keeping models up-to-date through retraining or fine-tuning introduces a compatibility challenge: operational data typically exists in formats unsuitable for LLM consumption, requiring transformations into structured representations before it can be used effectively.

This session demonstrates how to bridge that gap by using Debezium as the backbone of a Retrieval-Augmented Generation (RAG) pipeline. We’ll dive into the mechanics of capturing transactional changes from relational databases, converting text fields into vector embeddings via Debezium’s Embeddings SMT, and streaming them into a vector database for efficient retrieval.