Abstract
This disclosure describes a multi-tenant Configure-Price-Quote (CPQ) software architecture in which a Large Language Model (LLM) configures complex technical products through natural-language conversation while guaranteeing that every selected option is valid for the relevant product line. Validity is enforced not by translating the conversation into a formal constraint language and running a constraint solver (SAT/CSP/MIP/MiniZinc), but by restricting the LLM's action space: the LLM can only invoke a small, fixed set of tools that operate against a curated relational database, and it can only pass option identifiers that the database itself has previously returned. Hallucinated or invalid identifiers are structurally impossible because they fail at the database layer. The same approach applies to chat and voice (speech-to-text → same pipeline → sentence-anchored text-to-speech with mid-utterance tool dispatch). Multi-tenant isolation is achieved with a single shared configurator container, a reverse proxy that injects a tenant-slug header per ingress port, and one isolated database per tenant. This document is a defensive publication released into the public domain under CC0 1.0 Universal.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Hinterdorfer, Christian, "Natural-Language Multi-Tenant Product Configurator Using LLM Tool-Calling Against a Curated Relational Database, Without a Constraint Solver", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10033