Anonymized case
Corporate platform with multiple agents and knowledge bases
The project required agents with distinct roles, document access controls, and usage tracking for each operation.
Scenario
A platform where distinct departments need AI over their own policies, procedures, and files. One agent must not see the other department's folder. Documents live in storage. Model consumption needs to be visible to whoever pays the bill.
Problem
Putting “one company AI” into a single generic prompt. Without a binding between agent and knowledge base, the answer mixes commercial policy with internal procedure. Without document control, the index ages. Without a consumption report, the project becomes a surprise on the invoice.
Diagnosis
The design required purpose-specific agents, clearly scoped knowledge bases, role-based permissions, RAG for frequently updated documents, and a production environment for the index, storage, and API.
Solution
The implementation treated agents as an operational product, not as a demo.
- Several agents, each with its own purpose and system prompt.
- Document-based knowledge bases with clear indexing criteria.
- Permission: the agent only retrieves what that profile can see.
- Storage for files; the application references them, it does not mix binaries with the prompt.
- Consumption tracking and a path for “I don't know, I'll transfer.”
- Integration with the rest of the business system when the question was transactional data, not text.
Result
The company started having usable AI by function, with controlled documents, instead of a single chat with no owner. We do not publish internal department names or token volume. The result that can be stated is design: RAG with permission, separate agents, and production operations — not a generic assistant with a loose PDF.
Technologies used
- AI agents
- RAG and knowledge bases
- Object storage
- APIs and user permissions
- Consumption reports
Related services
AI, development, and infrastructure in the same product.