LangChain Jev Examples
Learn typed Jev judgments, model routing and tool gating through runnable LangChain examples.
Added to Jevfast
How it uses Jev
The course demonstrates typed Jev decisions in three LangChain patterns: classify a support ticket with Choice, Noul, and Score questions; route each task to a fast or powerful model with Choice; and gate a destructive tool call on whether the user requested it. Application code reads the typed answer to report classifications, choose a model, or allow/block the tool.
What Jev decides
Support-ticket classification example (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · department
YOUR APP
INSTRUCTION
Choose the team that should handle this ticket.
STATE
The source-defined ticket text describes a payment integration outage, prior unanswered emails, and a cancellation threat.
JEV · CHOICE
- billing
- technical
- sales
- other
Question 2 · urgent
YOUR APP
INSTRUCTION
Decide whether the ticket requires a response within the hour, using the source criteria about blocking or financial loss.
STATE
The source-defined ticket text describes a payment integration outage, prior unanswered emails, and a cancellation threat.
JEV · NOUL
YesNo
Question 3 · frustration
YOUR APP
INSTRUCTION
Score the customer's apparent frustration using the three ordered levels from calm to angry.
STATE
The source-defined ticket text describes a payment integration outage, prior unanswered emails, and a cancellation threat.
JEV · SCORE
0
App workflow
- Classify a ticket
Define Choice, Noul, and Score questions with criteria, pass ticket state and questions to `TypeSafeClassifier`, then read typed result maps.
- Route an agent task
Configure fast and powerful model choices, attach router middleware to an agent, and invoke it with task text.
- Gate tool calls
Register note tools and configure AutoMode to gate `delete_note`; invoke the agent, then inspect remaining notes and any blocked tool error.
Why it is interesting
The examples show how one probabilistic decision layer can serve different control points: extracting several ticket attributes in one request, choosing model cost/capability per task, and checking user intent before a destructive tool runs.