Agentic supply-chain chat
Answers come from tenant data, with citations. The model picks a query, the backend injects the tenant and runs it. Tools filtered by role. A 25-question eval gates every release.
Claude APITool use
The rule
The model never answers from memory. Every answer starts with a real query against the tenant’s data. If the data doesn’t have the answer, the chat says so.
The loop
- The user asks a question.
- The model picks one of a fixed set of query types.
- The backend injects the tenant id and runs the query. The model never touches the database.
- The model writes the answer from the results, with citations to the exact records.
Safety
- The tenant is always set server-side. A user can’t reach another tenant’s data, even by asking.
- Tools are filtered by role before the model sees them.
- Per-tenant daily limits keep cost in check.
Quality gate
A 25-question evaluation suite has to pass before any release. It checks query choice, no facts beyond the data, citations, and graceful “data not available” answers.