Design and Implementation of Bilibili's Intelligent Customer Service System

Since the operation of the activity platform, many ad‑hoc issues arise when operators use the platform, such as unclear component functions or unexpected activity behavior. Traditionally, a "fire‑fighting" group of more than a thousand developers would answer questions in a WeChat work group, and the weekly problems were manually recorded in Excel. This manual approach caused high labor cost, error‑prone data, and difficulty tracking messages.

To address these problems, an intelligent customer‑service system was built that can automatically converse with operators, pull groups with one click, and persist FAQs.

System Overview

The system consists of five main parts:

Conversation UI

Conversation state machine

Data source model

Statistics and reporting backend

Integration configuration

Conversation UI

Operators prefer using the native WeChat Work (企微) app for communication, so the system adopts a WeChat Work application account as the primary dialog entry. The UI is embedded in the app, avoiding third‑party web pages that operators dislike.

Message Reception & Parsing

All user actions in the WeChat Work app are captured via a callback interface. In the test environment a proxy was added to forward external requests to the UAT domain, and a department flag is attached to each message to keep data isolated across departments.

Conversation State Machine

The backend maintains a full conversation lifecycle: opened, in‑progress (human‑hand‑off or not), and closed. Each time an operator sends a message, a conversation is opened. Various message events trigger state transitions, FAQ matching, answer generation, one‑click group pulling, and automatic closure after inactivity. A delayed‑message queue handles long‑silence confirmations.

Data Source Model

Two models are used for answer generation:

Elasticsearch‑based search: the built‑in tokenizer and token filters split the user query and retrieve the most similar FAQ from the knowledge base.

Large‑language‑model‑based approach: an OpenAI‑powered vector similarity model (ChatGPT‑like) and a domain‑specific SimBERT model trained on collected dialogs. The system can switch between these models at runtime.

To keep the models up‑to‑date, an offline pipeline continuously feeds newly collected FAQs and conversation logs into the training jobs, allowing the models to improve over time.

Statistics & Reporting Backend

The backend is a low‑code visual management console built with the internal LEGO system. It provides:

Conversation detail view for review.

Remark and status tracking per conversation.

One‑click FAQ upload that feeds both Elasticsearch and the LLM pipelines.

Clonable pages so other teams can quickly spin up their own admin UI.

Integration Configuration

After the core system was completed, it was opened to other internal teams (frontend infra, DBA, etc.). Integration steps include providing WeChat Work app credentials, binding the unified callback, and cloning the low‑code admin page. This enables each team to have its own isolated conversation state machine and data source while sharing the same underlying platform.

Practice and Results

Since launch, the system has handled about 1,000 operator issues, averaging five daily queries. After adding ChatGPT‑based answering, the intelligent resolution rate increased by roughly 7%.

Future Outlook

Planned improvements include:

Providing an SDK so any internal team can embed the chatbot without custom development.

Automating the onboarding workflow with an online platform and approval mechanisms.

Enhancing the human‑hand‑off experience by working with WeChat Work to allow direct in‑app group creation, reducing the need for external group pulls.

Unexpected Benefits

The project stayed lightweight, using a domain‑driven microservice architecture and clear interface boundaries (MVA – Minimum Viable Architecture). This simplicity made the codebase easy to maintain while still supporting long‑term extensibility.

Overall, the system demonstrates how a well‑designed, AI‑enhanced service can reduce manual workload, improve response quality, and be reused across multiple product teams.