Skip to main content

Prompt Optimization

Define and automatically optimize LLM-based workflow — without writing code. Prompt Optimization lets you design flexible LLM workflows using a visual builder.

Step-by-Step Guide​

1. Select an Optimization Option​

After entering Prompt Optimization, choose between:

  • Free-form Optimization – Manually configure inputs/outputs and craft the logic.
  • Data-Driven Optimization automatically iterates and improves prompts using both pipeline structure and labelled dataset examples.

Once selected, a visual workflow builder editor will appear with two default blocks: Input and Output.

2.1 Free-form Optimization. Configure Input and Output Fields​

Click on the Input or Output nodes to set:

  • Data Type – string, integer, or float.
  • Field Name – Define a label for each field (e.g., user_query, answer).
  • Add more fields with "Add input/output field" if needed.

2.2 Data-Driven Optimization. Upload Dataset for Data-Driven Optimization​

If using a dataset:

File Requirements:​

  • File Size: Up to 50 MB
  • Columns: At least 2
  • Rows: Minimum 10 rows

You can either:

  • Upload your own .csv file
  • Or use the Sample Dataset, which includes two example columns:
    • phrase
    • translated formal phrase

Once the dataset is loaded, you can configure column behaviour using the Set type for each column tool.

Column Configuration Features:​

  • Use Find column to search by name
  • Define columns as input or output
  • Sort columns alphabetically (A–Z) or in reverse (Z–A)
  • Use Edit Columns to rename or modify columns
  • Use Export to download the modified dataset

Once you finish configuration, click Continue to move to the workflow editor where you can define the workflow logic and see your optimized prompt flow in action.

3. Add Logic Nodes: Gate and Processor​

Click the + icon (bottom-right toolbar) to add:

  • Gate

    • Provide a Hint – a guide used by the optimizer to generate the final gate logic.
      Example: Determine if the input data is in English or not.
    • Define Input data type and name.
    • Add Conditions to define under which input criteria this node is triggered.
      Example: Condition — "English text" (input is in English), "Not English text" (input is not in English).

  • Processor

    • Provide a Hint – Give a brief description of how the input should be processed.
    • Set Input and Output fields similar to Gate. It takes input from upstream nodes and transforms it according to the provided processor hint.

Connect nodes visually to define the workflow. Processors can be chained and combined Gate nodes for advanced flows.
Each node prepares its part of the final flow.

4. Set Up Model Provider​

Before running prompts, configure the model backend:

  1. Go to the Provider tab.
  2. Select the provider in the Model Provider dropdown (OpenAI supported, Ollama currently not available).
  3. Add your API key.
  4. (Optional) Allow storing the API key in your browser’s local storage for future sessions.

5. Define Task Optimization Settings​

Scroll to the Task Optimization section:

  • Task Description – A summary of what the workflow should achieve.

  • The Teacher Model prepares the optimization logic and generates the "ideal prompts" based on hints and (optionally) a provided dataset.

  • When ready, the Student Model is evaluated by running through the same workflow — using the optimized prompts to generate outputs. After the Student model completes its run, the results can optionally be reviewed again by the Teacher model or evaluated using scoring metrics. When the model is ready the model will be run by the Student model.

  • Evaluation Metrics:

    • Exact Match – Optimization continues until model output matches expected result exactly.
    • LLM-Based Scoring – Measures similarity with a provided criteria to target output (ideal when exact match is too strict).
    • None – Skip evaluation metrics.

6. Run Prompt Evaluation​

  1. Click Run the Model (top-right corner).
  2. Add input examples manually or upload a .csv file for batch testing.
  3. Click Predict to generate model outputs.

Quick Tips​

  • You can chain multiple gates and processors for advanced workflows.
  • Use smaller models (e.g., gpt-4o) for testing before scaling up to larger models to save costs.
  • Store your API key in local storage to avoid re-entering it each session.