One tenant per company
LiveAll data carries a company ID; WeCom credentials, model keys and MCP credentials are encrypted per company and invisible to others.
One language, one repository, one deployment. Browsers never touch the database directly; every business table has row-level security, so even an application bug cannot reach another company's data.
As of October 2026.
The WeCom long connection receives the @-message and finds the object bound to the extension.
The trusted session tells who, which company and which role; client parameters are never trusted.
The permission matrix is checked per action, deny by default.
The data layer queries with the user's identity while database row-level security applies.
The question is embedded, similar passages retrieved and added to the prompt.
When retrieval returns nothing, code enforces "do not make things up", not just the prompt.
The model streams its answer and the group message updates as it goes.
Call logs, token usage and audit records are written; a logging failure never blocks the reply.
Chat, streaming, tool calls and embeddings go through one module; Qwen by default, selectable per agent.
Configure OpenAI-compatible endpoints (e.g. DeepSeek, Zhipu, private vLLM) for chat, embedding, rerank, speech-to-text and text-to-speech; errors fall back to the platform default.
Every call records tokens, model and source; usage over the 30-day quota trips a circuit breaker.
On-premises servers hold only a revocable, metered token; real keys never land on them.
All data carries a company ID; WeCom credentials, model keys and MCP credentials are encrypted per company and invisible to others.
Membership decides which companies you can enter; the active company decides who you act as; switching goes through a controlled interface.
Open, suspend (immediately) and re-enable; each tenant has model and storage quotas.
Role × page, role × data scope and department × business data in one console.
| Aspect | SaaS | On-premises |
|---|---|---|
| Servers | AIPaddle cloud | Your own Ubuntu servers, alongside existing systems |
| Database | Shared, with row-level multi-tenant isolation | A dedicated database holding only your company |
| Scope | Everything, including platform operations | Platform operations removed at build time |
| Deliverables | — | Compiled output and configuration only; no source or internal documents |
| Models | Platform keys or your own | Through the model gateway, real keys stay off your servers; or bring your own |
| Releases | Merge → full tests → automatic deploy | Manual approval → build, trim, verify → migrate → atomic switch → automatic acceptance, rollback on failure |
| Backups | Platform backups | Daily backups kept for 14 days |
Only commits that passed the full test suite in staging.
All 7 required settings must be present on the server.
Platform operations features are removed.
Compiled on our build machine.
Passes only with no source, no internal text and no operations routes.
Copied to a new directory on the server without switching yet.
Database migrations run locally and stop on failure.
Atomic swap and restart.
Checks version, login page and that operations routes are unreachable; rolls back on failure.
Reversible actions run directly and can be undone within 12 hours; irreversible ones need the owner's approval.
Every save is scanned: hard-coded secrets blocked, injection risks fixed.
Positions are the work and drive assignment; roles are permissions: sponsor, owner, member, supervisor.
Four scopes: all, department and below, department, self.
All sharing and grants of AI capabilities are approved in one place, showing who gets access now and who has it already.
Work records and KPIs are visible only to the PM, sponsor and admins, never posted to groups; full chat logs are not collected.
A 30-minute online demo built around your industry, including how we would connect to your WeCom and ERP.