AI Deployment and Customization

Introducing Zhichao AI's privatized deployment, GPU computing power, dedicated agents, industry applications and on-demand development.

Private deployment method

Zhichao AI privatization can be divided into three categories based on data security, budget, computing power, model calling and network search requirements. Actual projects should be subject to formal implementation evaluation.

Method	Description	Applicable scenarios
All privatized	Deploy approximately 32B models such as DeepSeek and Qianwen on self-built servers, and model inference and knowledge base capabilities should be run within the customer environment as much as possible	Customers who have strict restrictions on data going out of the domain and have sufficient budget and GPU computing power
Vector localization	Deploy vector models and knowledge base storage operations on self-built servers, and use public network large model APIs for question and answer reasoning	Customers who want the knowledge base storage process to be locally controllable and accept external model reasoning
Vectors are only stored locally	The self-built server saves the vectorized knowledge base, and model inference and vector operations use the public network large model API	Customers who want to reduce local GPU costs and accept external model services

Reference server configuration

Deployment method	Server purpose	Reference configuration	Card counting configuration
All privatized	Computing server	CPU 28368Q, 256G memory, 960G SSD 2, 8TB HDD *2	5880 ADA 48G *4
Vector localization	Vector server	CPU 24310, 128G memory, 480G SSD, 4TB HDD 2	3090 24G *2
Vectors are only stored locally	Vector server	CPU 24310, 128G memory, 480G SSD, 4TB HDD 2	None

External models and network searches

Some deployment methods will use the public network large model API, Internet search API or TokenPlan. Optional models and services can include Minmax, GLM, DeepSeek, and Tencent Cloud Internet Search API. The cost will vary with the model, call volume, number of Internet searches and enterprise usage intensity. The official cost is based on the commercial quotation and implementation plan.

Deployment form

Zhichao AI can be used as an optional module of BabelBird Enterprise Drive, and can also be deployed according to enterprise requirements in privatization projects. Private deployment usually requires additional GPU computing power, model services, vector retrieval or knowledge base indexing services, and needs to be combined with the enterprise's data security, network access and third-party model policy design.

On-demand development

Babel can develop exclusive AI applications around enterprise industry scenarios and business processes, such as exclusive agents, business data queries, industry-specific Q&A, document processing processes, AI tags and classification rules, third-party system access, etc.

Typical Delivery Content

Data scope confirmation, knowledge base design and agent role design.
Configuration of capabilities such as models, vector libraries, OCR, multi-modality, and third-party data sources.
Permission linkage, audit logs, access scope and security policy design.
Agent embedding, enterprise portal integration or third-party website access.
Test question and answer set, effect optimization, online training and subsequent maintenance.

Privatization considerations

Whether the privatized Zhichao AI runs completely within the internal network, whether it calls external models, whether it uses third-party APIs, and whether it supports offline models should be clarified in the implementation plan. Customers involving sensitive data should prioritize data flow, log retention, model training boundaries, and permission audits.

Permissions and usage boundaries

Zhichao AI’s answers, searches, and file references should follow Babel’s existing permission system. Users can only access content within their account, department, project, share, and file access controls.
Materials used for learning, training, indexing, or Q&A should be specified by the enterprise or administrator; all files should not be considered public knowledge sources by default.
External customer service, website embedding, third-party data sources, privatized models and GPU computing power are optional deployment or customization capabilities and should be subject to actual authorization, implementation plan and enterprise configuration.
When AI answers involve high-risk content such as contracts, finance, medical, legal, and engineering safety, the AI output should be used as auxiliary information and reviewed by professionals.

Permission system
Security and Audit
FAQ: Deployment, AI and technical issues
Public introduction: <https://zhuanlan.zhihu.com/p/2052372550513072059>