Everyone’s talking about RAG (Retrieval-Augmented Generation), but let’s be honest: the standard demo usually goes something like this: split your content into chunks, toss them into a vector database, cross your fingers, and hope the LLM comes back with something vaguely useful. That’s fine for a project with straightforward content, but in the case of a hundred thousand Drupal nodes, you need a different RAG setup and strategy.

In this session, we’ll look at how to make RAG actually reliable. We’ll cover:

Introduction to RAG and embeddings: a tour through the basics of combining large language models with external knowledge.

Hybrid retrieval: mixing semantic embeddings with old-school keyword search (BM25) so you don’t miss the obvious.

Smarter chunking: using headings, paragraphs, and semantic cues instead of arbitrary splits.

Re-ranking: allowing smaller models to refine the retrieval process before the LLM takes over.

Context injection: using user profiles, session data, and relationships between nodes to give the model the “bigger picture.”

You’ll leave with a clear sense of how to move from basic “embeddings in a box” to a proper AI assistant that respects your content structure, scales with Drupal, and maybe, just maybe, makes your users think you’ve built magic.

Bring your curiosity, and perhaps a cup of tea. We’ll be doing more than chunking this time.