Sunshine Insurance Group's Zhèngyán Large Model Open Platform: Architecture, Tools, and Business Applications

General artificial intelligence is driving a technological revolution, with large models serving as a new engine that creates opportunities for the insurance industry, such as intelligent marketing, digital transformation, and ecosystem construction.

Sunshine Insurance Group has built the Zhèngyán Large Model Open Platform centered on GPT technology, linking multiple external large‑model services and, on the basis of private deployment of open‑source models, injecting Sunshine’s own knowledge and data to form a vertical insurance‑specific GPT foundation.

The platform consists of three layers—Tool Layer, Sunshine GPT Layer, and Business Application Layer—offering enterprise‑level Model‑as‑a‑Service (MaaS) capabilities.

The Tool Layer provides large‑model development utilities, including automatic training, evaluation, a Prompt factory for prompt management and optimization, and a unified plugin marketplace for dynamic development and management of plugins.

The Sunshine GPT Layer implements smart routing to dynamically schedule external and self‑developed models, applies intelligent review models for safety and compliance, and delivers professional, generic, and personalized insurance capabilities.

The Business Application Layer follows a “1+3+N” model: “1” empowers all office scenarios (text creation, summarization, image generation); “3” builds sales, management, and service robots with product explanation, diagnostic, and customer‑query abilities; “N” expands to additional scenarios such as precise product design, pricing, and automated reporting.

AutoTrain, the large‑model development tool, unifies data, model, and metric standards, allowing users to select an NLP task, upload data and configuration, and start training without complex environment setup, thereby improving development efficiency.

Sunshine’s self‑developed large model, selected after evaluating 12 open‑source Chinese models and fine‑tuned with knowledge injection, achieves a 15% increase in information‑extraction accuracy, 5% in intent‑recognition accuracy, and an 8.7% boost in intelligent‑QA answer rate compared with the baseline BERT model.

Smart routing provides a flexible model‑selection mechanism with three effective levels (interface, system, global) and both rule‑based and intelligent strategies to ensure the most suitable model handles each request.

The plugin marketplace, built on LangChain, connects the large model with external tools through an agent component, parses user intent, invokes appropriate plugins, and aggregates results for the final response.

Intelligent review is a multimodal content‑moderation system that uses keyword interception, similar‑sentence detection, and dedicated review models to ensure generated content complies with laws and regulations.

Seven key capabilities are delivered: professional (insurance‑specific semantic understanding, QA, system integration), generic (dialogue, multimodal, code), and personalized (Sunshine cultural lecturer and “Sunshine Sheng” sales persona), all powered by knowledge injection and knowledge mounting.

Professional capabilities are enhanced through knowledge injection and mounting, improving extraction, intent recognition, and QA accuracy; over 20,000 annotated data points support these improvements.

Personalized capabilities feature AI personas that provide brand‑aligned interactions, with a FAQ‑DocQA‑Chat pipeline that segments insurance documents, vectorizes them, retrieves relevant passages, and feeds them to the LLM for precise answers.

Three robots—sales, service, and management—are equipped with specific abilities (product presentation, emotional support, operational diagnostics) and are deployed across multiple channels (website, WeChat, app, call center) to achieve end‑to‑end insurance service.

Future plans include adding more AI personas, integrating the platform with the company’s dialogue system, and further advancing large‑model applications across insurance operations, risk management, and customer experience.