Dataset Optimization: Improve Prompts Directly in Your Dataset

If you already have a dataset with a Run Prompt column, skip to Step 2.

Go to app.futureagi.com → Dataset (left sidebar) → Add Dataset → create a dataset with input columns (e.g., question, context).

Add a Run Prompt dynamic column:

1. Click Add Column → select Run Prompt
2. Write a prompt template referencing your input columns; for example: Answer this question using the context: {{question}} Context: {{context}}
3. Select a model (e.g., gpt-4o-mini)
4. Run the prompt to generate outputs for all rows

The Run Prompt column stores the prompt template and generated outputs. This is what the optimizer will improve.

Navigate to your dataset → click the Optimization tab (fourth tab, after Data, Annotations, and Experiments; before Summary).

This tab shows all optimization runs for this dataset. If no runs exist yet, you see an empty state with a Run Optimization button. Once runs exist, the list view shows an Optimize Prompts button in the header.

Click Run Optimization (empty state) or Optimize Prompts (list view header) to open the configuration drawer.

Field	Value
Name	Auto-generated (e.g., Prompt-GEPA-Mar04-1430) — edit if needed
Choose Column	Select a Run Prompt column from the dropdown
Choose Optimizer	Select an optimization algorithm (see table below)
Language Model	The LLM used during optimization (e.g., gpt-4o)
Optimizer Config	Parameters specific to the selected optimizer (auto-populated with defaults)
Evaluations	Select one or more evaluation templates to score candidates

Available optimizers

Optimizer	Config parameters	Best for
Random Search	num_variations	Quick baseline — generates random prompt variants
Bayesian Search	min_examples, max_examples, n_trials	Few-shot example selection and ordering
ProTeGi	beam_size, num_gradients, errors_per_gradient, prompts_per_gradient, num_rounds	Targeted prompt edits based on error analysis
Meta-Prompt	task_description, num_rounds	General-purpose prompt rewriting
PromptWizard	mutate_rounds, refine_iterations, beam_size	Multi-stage mutation, scoring, and critique-refinement
GEPA	max_metric_calls	Evolutionary exploration of diverse prompt styles

Click Start Optimization to launch the run.

After launching, the Optimization tab shows the run with its current status:

Status	Meaning
Pending	Queued, waiting to start
Running	Actively optimizing — auto-refreshes every 5 seconds
Completed	All trials finished
Failed	An error occurred during optimization
Cancelled	You stopped the run manually

Click the run to see the detail view with:

• Steps: progress through the optimization stages
• Results graph: score progression across trials
• Trials grid: each trial’s score and prompt variant

Click any trial in the grid to open the trial detail view. The detail view has two tabs:

Prompt tab: shows a side-by-side comparison:

• AGENT PROMPT: the baseline prompt from your Run Prompt column
• OPTIMIZED AGENT PROMPT: the variant generated by the optimizer for this trial
• Toggle Show Diff to highlight changes between the two prompts

Trial Items tab: shows the individual iterations the optimizer ran to produce this trial’s prompt, with input/output text and evaluation scores per row.

Review multiple trials to see how different optimization paths produced different prompt structures. The best-scoring trial’s prompt is your candidate for promotion.

Once you’ve identified the best trial:

1. Copy the optimized prompt from the trial detail view
2. Update your Run Prompt column’s template with the improved version, or
3. Save it to the Prompt Workbench for version control and production serving

To re-run optimization with different settings (e.g., a different optimizer or metric), click Optimize Prompts again from the Optimization tab. Previous runs are preserved for comparison.