Evaluate
Ask several questions about one piece of content in a single step.
Type
typesafe.evaluateFolder
typesafeChannels
Any
Pipe
Value
Scope
Any
Evaluate sends every question in one step. They all read the same State, and each answer comes back under the Id you set. Mix Choice, Score, and Noul.
Ask every question you might need, including ones that only matter for some inputs. Later blocks can ignore answers they do not use. A question cannot see another question's answer. If the next question depends on an earlier answer, use a second block after this one.
Configure
| Setting | Description |
|---|---|
| Model | jev-latest (default), jev-preview, or jev-1.13.0 |
| State | The text or JSON every question reads |
| Id | Name for this answer. Later blocks read answers.<id>. The model does not see the id. |
| Type | Choice, Score, or Noul |
| Instructions | One specific question |
| Options | Shown for Choice. Key and description, 1 to 255. |
| Levels | Shown for Score. 2 to 10 descriptions, low to high. |
| Yes means / No means | Shown for Noul. Optional. |
Handles
| Handle | Role |
|---|---|
| In (target) | Incoming connection |
| Out (source) | Continues after every answer returns |
Pipe schema
The next block reads this on the data pipe (pipe.value).
| Key | Type | Description |
|---|---|---|
model | text | The Jev version that answered |
answers | object | One answer per id |
usage | object | input_tokens and output_tokens for the whole call |
Each answer has a type.
Choice: choice, probabilities, confidence.
Score: score, legend, probabilities, confidence.
Noul: noul. Evaluate always returns the probability. Yes at is only on the Noul block.
Use case
- Classify the team, score frustration, and check urgency on one ticket, then branch on the fields you care about.