A bilingual (English / 中文) restaurant booking agent that never books until the guest says yes, and never fills in a value the guest didn't give.
Rules, not a model, read the date, time, party size and name. Anything unclear becomes a question. The only action, create_reservation, runs after an explicit yes.
Nalu is a restaurant group's guest assistant for questions and table bookings in English and Chinese, now in pre-launch testing. The booking path makes no model call at all; the model only handles what's left, and it can never write a booking field.
lisa@example.com became the name "Lisa", "what time do you close?" became the time "close", "I want to book…" became the name "I".| Model-only pilot (qwen2.5:7b) | 17 of 30 messages fully correct (57%); 3 invented values |
| Rules + confirmation gate, untuned held-out set | 33 of 39 fully correct (85%), 0 wrong or invented values; all 6 misses were omissions, so the guest was simply asked again |
| Development set | 35 of 35, gated in CI on every change to the agent. It was tuned against, so it is a regression suite, not an accuracy claim |
| Live replay | 23 conversations × 3 runs, 114 turns, 0 failures; p95 1.27 s on a later 38-turn replay |
| Local GPU | NVIDIA DGX Spark (GB10): about 47 tokens/s warm, about 10 s cold start, concurrent requests served one after another. Hence at most one model call per turn, none on yes / no |
| Honest limit | 54 non-booking conversations with zero wrong bookings only bound the rate below about 5.5% (rule of three). A safety claim needs 100+ adversarial trials |
The first version let a language model read the booking details. Measured, it filled in values the guest never gave, so the model came out of the booking path. If you run Ollama, this page can ask your local model to read the same conversation and show where it would have guessed. Nothing leaves your computer.
What is inside that model: architecture card for Qwen2.5-7B-Instruct. 7.62B parameters over 28 layers, grouped-query attention (28 heads, 4 KV heads), 14.2 GiB of weights in bf16 and 56 KiB of KV cache per token, so 32K tokens of context cost 1.75 GiB and the GPU budget goes to weights, not conversation history. Generated from the config alone, no weights downloaded.
# once, then restart Ollama
ollama pull qwen2.5:7b
OLLAMA_ORIGINS="https://kaihuan-huang.github.io" ollama serve
Chrome may ask to let this page reach your local network; allow it. The model reads only your messages in the chat above.
qwen2.5:7b on Ollama, temperature 0, 20 conversations, 25 Sep 2026. In 2 it chose AM or PM on its own when the guest hadn't said; in 1 it put "下周五" (next Friday) on a Tuesday. The rules asked instead, or got the date right.
Every bug fix starts as a failing test. These synthetic conversations run in your browser against the code on this page, with a fixed clock (Fri 25 Sep 2026, 10:00). Each one also checks that no reply promises an email, text or call.