THIS is how you build agentic RAG systems that work

In this video from the Weaviate channel, Victoria argues that agentic RAG pipelines usually fail because of the context they are given, more than the wording of the prompt. She splits a working pipeline into five systems and takes them one at a time. The video runs a little over 11 minutes.

After the intro, the chapter at 0:46 makes the case that longer contexts are not better. At 1:18 she lays out the context engineering system as a whole before starting on its parts.

Query augmentation comes first, at 1:36: turning messy input from a person into something the pipeline can use. A chapter on query agents follows at 2:25.

Retrieval and chunking run from 3:25. The description says chunking strategy alone can make or break RAG performance, and it links Weaviate's chunking playground for trying different strategies.

Memory starts at 5:42. According to the description, this part contrasts short-term and long-term memory and explains why dumping the conversation history into the context is a failure mode. The description also links Weaviate's Engram product; the sources do not say how much of the chapter is about it.

Tools get their own chapter at 8:02. The point made here is that most tool failures are really context engineering failures.

Agents, the fifth system, come in at 10:08. The description calls them the orchestration layer that ties the other four together, and the place where every context failure compounds. A short conclusion follows at 10:51.

The description links Weaviate's context engineering ebook for a longer written version of the topic. The video is a fast map of the parts, suited to developers planning or debugging a RAG system who want a checklist of where context goes wrong. It comes from a vector database company, and all three linked resources are Weaviate's own.