The browser RAG (@crema/lexical-rag-ui) and the server RAG
(arcadia-search) coexist in this app, and the agent picks between
them via tool descriptions. The two libs each have their own README,
but neither covers the picking decision or how they relate. This doc
fills that gap.
- At-a-glance table: lib, engine, runtime location, corpus size,
update cadence, auth, agent tool, what it's best for.
- The system-prompt snippet that drives tool picking, with notes on
observed DeepSeek V3 behavior.
- Per-surface deep dive (build path, tools, storage, limits, why it
exists).
- Why both run by default (always-on fallback, A/B regression).
- Decision checklist for adding new corpora.
- "What lives where" cheat sheet pointing at the right file in the
right repo for common tasks.
Cross-references arcadia-search's README.md, MULTI_TENANT.md, and
ARCADIA_INTEGRATION.md so a reader landing here can navigate to the
right rabbit hole.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
