Building and Applying NIO's Enterprise Knowledge Platform: Architecture, Challenges, and Future Directions

Background

NIO historically operated multiple isolated knowledge‑base systems, creating data silos that made knowledge search difficult and raised concerns about completeness and accuracy. To address this, a company‑level knowledge platform was initiated with goals of integrating assets, providing a flexible management tool, establishing governance standards, enabling efficient knowledge promotion, and supporting large‑model corpora.

Overall Architecture Design

The platform is organized into layered components. At the lowest "raw material" layer, original corpora from various internal systems (documents, PDFs, code, etc.) are ingested. A unified knowledge production and operation tool then processes these materials—editing, reviewing, transforming, and analyzing—to produce structured knowledge assets. Assets are classified by confidentiality into public (structured and unstructured) and internal knowledge covering training, R&D, sales, quality, and manufacturing.

Knowledge services built on this foundation include content retrieval (full‑text, vector, catalog, and priority push) and rich‑text viewing, as well as interactive features. The platform supports downstream AI applications such as NIO APP, Nomi, and customer‑service systems.

RAG‑Based Intelligent Retrieval

For each document update, content extraction and slicing are performed, followed by vectorization and storage in a vector database. During query time, the user’s question is rewritten, vectorized, matched against stored slices, re‑ranked, and finally answered using a company‑specific large model with prompt engineering, returning both answer text and source citations.

This RAG workflow reduces knowledge‑search time by an estimated 40% compared with traditional manual search.

Challenges and Solutions

1. Answer Accuracy : The system categorizes queries into professional, sensitive, and ordinary. Professional and sensitive questions use keyword search to ensure precise, safe answers, while ordinary questions leverage vector search with risk warnings or fallback to the large model when necessary.

2. Permission Control : Role‑Based Access Control (RBAC) assigns roles and permissions, ensuring that users only retrieve knowledge they are authorized to see, preventing data leakage.

3. Multi‑Domain Smart QA : Separate knowledge bases per domain (R&D, production, sales, etc.) are queried according to the asker’s context, with answer style adjusted to fit the domain.

4. High Concurrency : A knowledge‑answer cache stores question‑answer pairs, updated in real‑time when knowledge changes. Three serving modes—No Cache, Only Cache, and Cache First—balance coverage and latency, with Cache First recommended for most scenarios, achieving >30% throughput improvement.

Future Outlook

Planned enhancements include AI‑assisted knowledge understanding (auto‑summaries, term explanations), AI‑assisted knowledge creation (drafting and expanding documents), multimodal capabilities (image‑based queries and mixed media interactions), and expanding supported knowledge formats beyond PDFs and web pages.

The overall vision is to continuously refine the platform, leveraging AI to improve knowledge consumption, creation, and accessibility across the enterprise.

Q&A

Q1: For professional or sensitive questions without LLM inference, does the system fall back to full‑text search? A: Yes, it uses full‑text retrieval to ensure accuracy and safety.

Q2: How does the cache maintain answer correctness without constant expert labeling? A: Answers remain correct as long as the underlying question, knowledge base, and model stay unchanged; updates to knowledge trigger automatic cache refreshes.