OPPO XiaoBu Assistant: Building a Low‑Code, End‑to‑End Dialogue System Platform

OPPO's XiaoBu Assistant is an integrated, multi‑type dialogue system embedded in smartphones and IoT devices, designed to support task‑oriented, knowledge‑based, and chat interactions.

The platform adopts a low‑code approach, allowing users to create, configure, train, and automatically deploy skills with minimal coding, while maintaining extensibility for diverse business scenarios.

Four key capabilities are highlighted:

Business domain modeling and a universal capability map.

Exploration of semantic understanding, supporting model‑based, rule‑based, and knowledge‑based NLU pipelines.

Componentized core functions organized as operators for preprocessing, retrieval, model inference, and rule execution.

An end‑to‑end offline platform that visualizes the entire skill lifecycle.

NLU is divided into preprocessing, front‑processing, intent detection, and slot extraction, with three strategies: model training, high‑precision matching, and knowledge‑driven recognition. Slot extraction combines dictionary‑based, generic, and third‑party models.

Intent recognition follows a four‑step workflow (standardized data definition, offline data augmentation, model training with knowledge distillation, and inference via local or TF‑Serving). Retrieval‑based intent detection leverages both semantic vector and text indexes, enriched with slot and synonym features.

Core services are decomposed into reusable operators (text engine, retrieval engine, model engine, rule engine) and orchestrated via DAG graphs, enabling scalable processing of millions of dictionary entries.

Large‑scale text processing splits workloads by hash‑based slots, allowing automatic scaling and re‑sharding as skill data grows.

Vector retrieval is implemented with a graph‑based HNSW algorithm packaged as an Elasticsearch plugin, providing clustered semantic and text search.

Dialogue management consists of three layers: strategy aggregation, execution collection, and result ranking, supporting weak and strong multi‑turn flows, customizable state machines, and online programming interfaces that embed Python in Java.

The platform defines a unified response protocol for text, audio, rich media, and quick‑app replies, using a template syntax tree to generate client‑specific messages.

An end‑to‑end offline pipeline ensures skill lifecycle management through cross‑environment data synchronization, staged testing (dev → evaluation → internal test → production), and continuous online data mining to identify and correct bad cases, feeding back labeled samples for model improvement.

Overall, the XiaoBu Assistant platform demonstrates a comprehensive, low‑code, AI‑driven solution for building, deploying, and optimizing conversational skills at scale.