Retrieval UX playbook for LLM features

Building retrieval-augmented features is less about the model and more about how teams see freshness, provenance, and gaps. This is the playbook we use when we ship a new RAG-backed experience.

Start with the surfacing pattern

We pick the interaction first, not the stack. Is this a sidebar helper that answers questions, an inline autocomplete, or a bulk report generator? Each shape changes the guardrails:

• Sidebar helpers: show sources and last-updated stamps on every message.
• Autocomplete: bias toward recall, then let the editor surface confidence.
• Bulk reports: show row-level provenance and let the reviewer reject and resubmit sections.

We sketch two failure states per pattern - missing data and stale data - and make sure the UI has a visible affordance for each.

Keep a narrow document contract

Instead of indexing “everything”, we define a contract for each document type: title, short description, owners, freshness, and labels. The labels keep evals sharp.