2026-09-23 — AI hotel guide reliability: found and fixed a real prompt-bloat bug

Brief mark, not a full changelog — see Concierge — Guest Chat, Translation & AI Hotel Guide for the full investigation. Three repos, all pushed to staging: CoreAPI, cdn-ai (a different team's repo — changes here are additive/ opt-in, existing callers unaffected), ConciergeGuestApp.

Root cause found - The AI hotel guide's honest-fallback message was firing for questions the seeded corpus plainly covered. cdn-ai's own audit.jsonl showed "Context size has been exceeded" and 140-180s timeouts, not a grounding gap. - Measured, not guessed: every /chat call (translation and the guide alike) was paying for cdn-ai's CDN-analytics SYSTEM_PROMPT (~3,268 tokens) plus its full tool schema list (~10k tokens) — 85-90% of a typical ~15k-token guide prompt, regardless of relevance.

Fixed — cdn-ai - mcp_client.py's tool-list fetch was opening a brand-new MCP session (fresh connection + handshake) on every single /chat call, even though tool schemas never change at runtime. Now cached, 10-minute TTL. - ChatRequest gained use_tools/system_prompt (both default to original behavior — no existing caller affected). A caller with nothing to do with CDN analytics can now skip both entirely.

Fixed — CoreAPI - domain/ai.Client sends use_tools: false + a short minimalSystemPrompt on every call — cuts a typical guide prompt from ~15k tokens to ~2.2-2.5k. - Ask() (the guide's synchronous answer) now runs on its own GuideTimeout (25s default), separate from translation's intentionally short RequestTimeout (6s) — the two were incorrectly sharing one budget tuned for "fail fast, don't stall a chat send," which is wrong for an answer the guest is already waiting on with a typing indicator. - Deepened the RAG corpus: 6 new policy topics (wifi, gym/pool, laundry, airport transfer, loyalty programme, local attractions), 3 existing ones enriched (pets, parking, cancellation) — token-budget-checked against the fix above; adds well under 1k tokens. - AI guide prompt now allows light markdown (**bold** for a key fact, - bullets only for 3+ items) — was explicitly "no markdown" before.

Fixed — ConciergeGuestApp - guideMarkdown.jsx renders the two patterns above, scoped to AI-guide bubbles only — department/staff chat and translated text unaffected. - Suggestion pills: 2 more (from the new topics), translated into all 11 supported languages; fixed a real bug where tapping a translated pill sent the English original as the guest's own message (visually broken — now sends whatever's displayed). - One real, verified Unsplash photo (hotel lobby) as a hero banner at the top of the AI guide thread.

Not done, deliberately deferred - Smarter retrieval (only send relevant corpus entries, not the whole thing) — full-context stuffing is still fine token-wise post-fix; swapping retrieval strategy this close to a demo/submission deadline was judged riskier than the benefit. - Auto-detect language reliability itself (as opposed to the already-shipped explicit-selector mitigation) — a small-model limitation, not something to chase this late.

See: Concierge — Guest Chat, Translation & AI Hotel Guide, AI — CDN-AI system