Copper Larder
A public demo chatbot serving as the greeter front desk for a British pub. It uses a multi-stage pipeline chatbot architecture in which a large model is the final step, preceded by lightweight checks to control cost and improve safety.
Timeline (last 90 days)
Copper Larder is a public chatbot demo for a British gastropub that answers menu questions, takes callback requests, and uses streaming output for replies.
Copper Larder's request pipeline contains 8 stages, of which only stage 7 consumes tokens.
Copper Larder sets up four lightweight gates before the model: session caps, complaint escalation to a human, script blocking, and exact-match caching.
Copper Larder has about 19 pure pattern-matching interception rules for high-frequency questions, at zero token cost, with answers coming from structured restaurant data rather than hand-written text.
Copper Larder stores dietary constraints (such as vegetarian) as session-wide locks in session state, injects them every turn, and filters menu cards in code, rather than trusting the model to remember them on its own.
Copper Larder's key output guardrail is preventing the model from confirming availability; if a reply promises that a table is free, the guardrail rewrites the reply and blocks it from entering the cache.
When the model is down, rate-limited, or the key is missing, every failure path in Copper Larder returns a gentle, on-brand message and a callback card.
Copper Larder streams replies via Server-Sent Events and pairs this with an aria live region so screen readers announce the text as it arrives.
The Copper Larder application does not store raw IP addresses, so there is no sensitive data to leak.
All Facts (9)
Copper Larder is a public chatbot demo for a British gastropub that answers menu questions, takes callback requests, and uses streaming output for replies.
50%UnverifiedCopper Larder's request pipeline contains 8 stages, of which only stage 7 consumes tokens.
50%UnverifiedCopper Larder sets up four lightweight gates before the model: session caps, complaint escalation to a human, script blocking, and exact-match caching.
50%UnverifiedCopper Larder has about 19 pure pattern-matching interception rules for high-frequency questions, at zero token cost, with answers coming from structured restaurant data rather than hand-written text.
50%UnverifiedCopper Larder stores dietary constraints (such as vegetarian) as session-wide locks in session state, injects them every turn, and filters menu cards in code, rather than trusting the model to remember them on its own.
50%UnverifiedCopper Larder's key output guardrail is preventing the model from confirming availability; if a reply promises that a table is free, the guardrail rewrites the reply and blocks it from entering the cache.
50%UnverifiedWhen the model is down, rate-limited, or the key is missing, every failure path in Copper Larder returns a gentle, on-brand message and a callback card.
50%UnverifiedCopper Larder streams replies via Server-Sent Events and pairs this with an aria live region so screen readers announce the text as it arrives.
50%UnverifiedThe Copper Larder application does not store raw IP addresses, so there is no sensitive data to leak.
50%