Interpreters

This page explains Curator Interpreters for the person who designs the process — analyst, coordinator, business owner. You do not need to know how to program. Whenever a screen name appears, it is translated right there, with an everyday analogy. The animations play on their own: hover to pause and click to watch again.

What it is

Think of a cooking recipe. You gather the ingredients, follow the preparation in stages, and at the end you have a finished dish. In Curator, that recipe is called an interpreter.

You hand over a piece of information. It goes through one or more reading stages, one after another. At the end, you receive a result ready to use — in an automation, in a test, or in AI Chat.

The idea is simple: separate the job of understanding a piece of information from the rest of the process. Instead of hiding the request to artificial intelligence inside each automation, you concentrate that knowledge in a named recipe with history and measured cost.

• Read an official letter and extract name, document number, and deadline.
• Classify a customer message as praise, complaint, or question.
• Summarize a long contract into three objective points.

When to use

Use it when the stage involves understanding language or images — not just applying a rule that never changes.

• The input text changes format from case to case.
• The rule is easy to explain in plain language, but hard to write as a rigid condition.
• The same reading repeats across several company processes.

The pieces of the recipe

There are five ideas. It is worth knowing each one before opening the screen.

Name	In plain language
Recipe (interpreter)	The complete package: name, steps, and what it needs to receive.
Step	One reading stage. It has a title, a written request, and the artificial intelligence model that runs it.
Input	What you provide when using it. It can be text or an image. On screen it appears as input.
Request	The plain-language text that tells the artificial intelligence what to do. The screen calls this an instruction.
Result	The output of the last step. Always text. The format is whatever you asked for in the request.

How information moves

When you open a recipe, the center of the screen is a left-to-right diagram. Each numbered box is a step. The arrows show the order. A step can reuse another step’s result — that draws an arc above the flow.

On the real screen you see the same idea as the animation: boxes with the step name and chosen model, tags for what each stage receives, and the arc when a result is reused.

In each step you set four things: a short name (e.g. “Extract tax ID”), the artificial intelligence model, what goes in to that step, and the request. The source of what goes in can be you (at run time), the previous step’s result, or the result of a specific earlier step.

The request you write

The request is plain-language text. On screen the field is called an instruction. A well-written request usually has four blocks:

1. Role and objective. Who the artificial intelligence is in that moment.
2. Answer format. E.g. “one field per line” or “a single word only”.
3. Short rules. What to do if the data is missing, what to avoid.
4. The content. The place where Curator pastes the real information.

That place in the text looks like this: {{oficio}}. It is the slot in the request. If the information is named oficio, you put {{oficio}} at the exact point where the text should go. Below the field, the screen shows name tags — click to insert the correct format.

Test and follow along

On the diagram, the run button opens the window for what goes into the recipe. You declare what the recipe requires, paste an example value, and click Run in flow. Examples are kept for the next test.

During the run, each box goes through Processing, Completed, or Failed. When the round finishes, click each step’s result to see that stage’s text. That way you find the exact point where information was lost — without guessing from the final answer alone.

Under Monitor in the recipe list sits the history: date, success or failure, which recipe ran, and estimated cost. The provider charges by amount of text, measured in tokens — pieces of a word. The number in Curator is an estimate for comparing and projecting scale; real billing is directly with the provider, using your access key.

Where to use it in Curator

In the Interpreters module you create, test, and maintain the recipe — and anyone on the team can run a one-off case. In AI Chat you can ask to list, create, adjust, or run a recipe through the conversation. In automations, the robot handles the clicks and delegates understanding to the recipe.

Tips

• One step, one task. If the request asks for three things, it is usually three steps. Short stages make fewer mistakes and are easier to fix.
• Ask for the exact format. “Answer with a single word only” or “one field per line, no explanation”. Without that, artificial intelligence tends to chat too much.
• Define what to do if the data is missing. Instruct a fixed value such as (not_found). Without a rule, artificial intelligence tends to invent.
• Test with ugly cases. The pretty document always works. Include the crooked scan, the truncated text, the atypical case.
• Do not forget {{name}} in the request. Without the slot, registered information never reaches the artificial intelligence.
• Give business names. “Customer complaint triage” is better than “Interpreter 2”. The name is what the whole team will see.
• Measure before you scale. Three good tests do not guarantee a hundred cases. Use Benchmarks to score accuracy in a batch and compare models with cost and speed.