The Golden Fork

A menu is a small, dense corpus: 22 items, every one a legitimate candidate. So retrieval is deliberately greedy — top-20 at a similarity threshold of just 0.30 — because the LLM reading two irrelevant dishes is cheap, but missing the one gluten-free pasta is a failed answer. The division of labor is explicit: Pinecone provides recall, the LLM provides precision — the retriever casts a wide net, and the model does the fine filtering by actually reading ingredient lists.

"No beef or pork" is an exclusion — and vector similarity is terrible at absence: embedding that phrase mostly finds beef and pork dishes. Exclusion and mood queries are detected and expanded into positive sub-queries ("chicken dishes poultry", "seafood fish salmon") that run against Pinecone in parallel and merge, deduplicated. Dietary keywords auto-apply metadata filters at the vector level; free-form ingredient exclusions are deliberately left to the LLM, whose golden rule is "never say 'I don't see it' if you can reason through the menu."

The order state machine is the backbone: pending → confirmed → preparing → ready → served → paid, enforced server-side by a transition table. The KDS can't skip steps, the webhook can't double-confirm, and every transition timestamps itself — which is what makes the elapsed-time badges and the live guest timeline possible.

The 50-case black-box regression suite — allergies, exclusions, moods, Cantonese, contradictions, off-topic traps — crashed on line one: `__dirname` is not defined, because the repo had moved to ESM after the suite was written, and nobody had run it since. A test suite that can't run is documentation, not protection. One fileURLToPath fix later it ran — and immediately earned its keep.

Four failures shared a symptom: the assistant answered a recommendation question with a clarifying question — polite, but against the product's own always-recommend rule. The trail led to retrieval, not generation: "no dairy, no gluten, no nuts" ANDs three filters and returned zero items, and the empty-menu prompt branch told the model to ask. The fix respects the design philosophy — on empty filtered retrieval, retry once unfiltered and let the LLM reason over the whole set. 50/50. The suite's last three cases then failed with HTTP 429: the product's own rate limiter, correctly throttling its own QA.

The camera found the best bug of the session. The customization modal centers itself with translateY(-50%), and its borrowed entrance animation ends with fill-mode: forwards — which permanently replaced the centering transform, pinning the modal's top edge to mid-screen. On a 1280×800 laptop, the confirm button rendered below the viewport in a fixed layer scrolling can't reach: every dish with a required customization couldn't be ordered at all. One dedicated modalIn keyframe fixed it, on camera.

The 50-case regression suite passed 50/50 while a showstopper modal bug sat in plain sight — because tests assert API responses, and the bug lived in CSS animation semantics at a specific viewport height. Recording demos at laptop resolution found it in minutes. Anything user-facing deserves one pass of watching the actual pixels.

Most RAG guidance assumes thousands of documents and fights hallucination with strict thresholds. A 22-item menu inverts the economics: recall is nearly free, precision comes from the model reading ingredients, and the real failure mode is over-filtering — exactly where the bugs were found. Tune retrieval to corpus size, not to habit.

The empty-retrieval branch — ask a clarifying question — felt like sensible defensive coding, and it silently violated the product's core promise in 8% of the regression suite. If a fallback can speak to users, it needs the same test coverage as the happy path.