Abstract
Automated conversational agents depend on real-time intent classification and entity extraction. Rule-based and shallow classifiers are brittle and costly to maintain, while unconstrained large language models (LLMs) add latency and token cost and can return output that departs from the expected schema. This disclosure describes a technique in which the two operations are handled by two services. In a synchronous service, a user utterance is converted into a dense embedding and matched by approximate nearest neighbor (ANN) search against embedded training phrases, and matched intents are resolved from a document database. Entities are extracted only when requested, by an LLM whose decoding is constrained to a JSON schema. In an asynchronous service, intent updates uploaded as comma-separated values (CSV) files trigger batch embedding, streaming vector index updates, and metadata synchronization. Low-latency, schema-conformant results are provided, and the intent catalog is updated without downtime.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Nigro, Leonardo, "Hybrid Vector-Based Intent Classification and Schema-Constrained Generative Entity Extraction with Event-Driven Intent Catalog Synchronization", Technical Disclosure Commons, (October 28, 2026)
https://www.tdcommons.org/dpubs_series/12510