How to Turn Generic AI into a Reliable Enterprise Code Assistant with Local Knowledge Bases

When AI assistants like ChatGPT or Copilot generate code that frequently hallucinates—producing non‑existent or non‑compliant code—developers face frustration. The ZTO Tech team tackled this by combining a local knowledge base, structured prompts, and an MCP service, teaching a general large model the company’s internal technical specifications.

Root Causes

General models know nothing about a company’s unique frameworks, component libraries, API specifications, or directory structures, resulting in code that cannot run or violates security and best‑practice guidelines.

Common Enterprise Development Challenges

Private tech stack: custom frameworks and internal component libraries such as zmi and zui.

Specific standards: strict coding style, directory layout, API usage, and security requirements.

Dynamic knowledge: internal API docs and business logic that the model cannot fetch in real time.

Core Solution Framework

The team built a three‑part system:

Store all internal documentation (framework docs, component APIs, standards, exemplary code) in a local vector‑search knowledge base.

Define the AI’s role, capabilities, and workflow with structured prompts (e.g., “ZTO front‑end coding expert”).

Use an MCP service to retrieve real‑time internal API schemas and enforce compliance during generation.

Benefits include immediate access to enterprise‑specific knowledge, ensuring generated code adheres to internal standards, reducing review cycles, and accelerating development.

Technical Stack

Claude model (120K context, strong code ability) integrated with the Cursor editor, which provides the local vector search engine without server overhead.

Key Outcomes

Significant improvement in code compliance—generated code meets standards on first pass.

Qualitative boost in development efficiency—developers focus on core business logic.

Accelerated onboarding—AI acts as an always‑available mentor.

Automated knowledge consolidation—building the local knowledge base structures the company’s technical assets.

The approach is generic and can be adapted to different tech stacks and business scenarios, representing a practical method for aligning AI assistants with enterprise engineering practices.