paieralabs/paiera-evidence-pack-demo

PAiERA Evidence Pack is a sanitized, reproducible demo that shows one narrow slice of a larger assistant system called PAiERA. It demonstrates a structured fact model for people and their food restrictions, a deterministic matcher that checks whether a food item conflicts with someone's allergy, logic for keeping people's information separate across multiple conversation turns, and a prompt assembler that turns all of this into text ready to send to a language model. The full PAiERA system, including voice, chat routing, the frontend, file uploads, and unrelated database tables, is not included here. The project runs entirely offline. Installing it requires Node.js version 18 or higher and a C compiler to build a small database library called better-sqlite3. No internet connection and no language model call is needed to install, run the demo, or run its automated tests. The demo uses three fictional people: Alice, who owns the account and has a shellfish allergy, her spouse Bob, who has no restrictions, and their child Clara, who has a peanut allergy. Three example requests show how the system behaves. Asking to buy a peanut cake for Clara triggers a confirmed conflict warning. Asking to buy an unspecified cake returns an unknown result rather than guessing it is safe, since a plain cake could still contain peanuts. Asking to buy fish for Clara shows that Alice's own shellfish allergy is never mistakenly applied to Clara. A key design rule is that the underlying model can only read the assembled prompt. It cannot add people, record allergies, or change any stored fact. All such changes happen only through a local seed script that runs before the demo. The matcher currently understands just three allergen categories: peanut, fish, and shellfish. Every other restriction is treated as unknown on purpose, since a broader list could not be checked with certainty. The project only supports English text, and it does not detect requests to cancel or forget a previous action.