Question 1

What actually happens when the model is unsure?

Accepted Answer

Three outcomes, and nothing runs quietly. At or above the threshold the dispatcher invokes the action and returns EXECUTED. Between the minimum and the auto-confidence value it returns NEEDS_CONFIRMATION with a pendingId, which the workflow module keeps as a suspended, persisted workflow execution you confirm or reject later. Below min-confidence it returns REJECTED with an AmbiguousIntentException. The score used is the weakest link across the action choice and every argument.

Question 2

Can the model invent an argument value?

Accepted Answer

No. A free-text argument is always one of the spans that were offered to the model, and a value the model made up is refused, because the model never writes an argument. Every resolved argument is then re-checked for presence, enum membership, type, length, control characters and slashes before anything runs, and checked again when a confirmed call finally executes.

Question 3

What breaks if my tinystruct version is older?

Accepted Answer

The routing pipeline needs the framework change in 1.7.34, where @Action(arguments = ...) metadata keeps order, the optional flag and the parameter's Java type, and string arguments convert to a Set or List of enums. That metadata is what the question generator reads. There is also a classpath gotcha: bin/dispatcher only puts target/classes, lib/*.jar and the tinystruct jar on the path, so every module your app uses has to be packaged into lib/ or you get a NoClassDefFoundError for the sibling client module.

Question 4

Is this just a wrapper around a chat completion?

Accepted Answer

The README is explicit that it is not a chatbot: there is no prompt API and no chat abstraction. Jev answers typed questions by returning a probability distribution over options the action declares, so the only things an action receives are enum constants, booleans and verbatim spans of the user's own input. The description you write in @Action is what the model reads, which is why it recommends saying what the action does and what it does not.

Question 5

What does it actually do?

Accepted Answer

It lets natural language invoke existing tinystruct @Action methods, using TypeSafe Jev as a semantic dispatcher. Jev answers typed questions by returning a probability distribution over options you supply, and it never generates text or values, so every argument is an enum constant, a boolean, or a verbatim span of the user's input.

Question 6

How are parameters turned into questions?

Accepted Answer

An enum becomes one choice question over its constants, a boolean becomes one noul question, a Set or List of enums becomes one noul per constant, and String, number or Date parameters become a choice over spans of the input with a not stated option.

Question 7

What are the key configuration settings?

Accepted Answer

The allowlist typesafe.routing.allowed-actions starts empty and nothing is routable until it is set, min-confidence defaults to 0.80, auto-confidence is off, and the strategy is single or two-stage. Caching can be none, memory or redis, and there is a separate confirmation service and workflow repository.

Tinystruct TypeSafe SDK

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community