RAG without faithfulness is search

RAG is often sold as the antidote to hallucination. Retrieval is necessary. It is not sufficient.

A system can retrieve the right chunk and still answer from the model's prior. It can cite a document it did not use. It can stitch two sources into a claim neither contains.

If you ship RAG without evaluation, you have search with a confident narrator.

The minimum contract I want:

• Context relevance — did we fetch something that could answer the question?
• Faithfulness — does the answer stay inside the retrieved evidence?
• Citation checks — do the pointers resolve, and do they actually support the sentence?
• Session memory — can a follow-up retrieve prior findings instead of starting over?

In AgentEval, RAG mode is a first-class suite, not a blog post. In Agentic Data Analyst, every specialist write-up is indexed in ChromaDB so later questions are grounded in the session, not in vibes.

The model is allowed to be fluent. It is not allowed to be unaccountable.