Promptuna

promptuna evaluates and optimizes functions that use an LM to accomplish a goal.

Such functions (hereinafter referred to as programs) are not thin wrappers around a single complete() call. Each program makes exactly one LM completion, wrapped in a deterministic scaffold — code that shapes inputs and renders the template before the call, and parses or repairs the model output after. In production, users rarely hit the raw completion; they hit the completion plus its scaffold. The harness evaluates that full product.

In the refinement loop below, promptuna provides the primitives for you to define the metrics that judge how well your program performs (3). Then, it can use those scores to drive automated improvements on the prompt template (4).

flowchart LR
    A[1. Make a program] --> B[2. Run the program]
    B --> C[3. Evaluate the program]
    C --> D[4. Improve the program]
    D --> B

The loop above maps directly onto the package layout:

promptuna.report sits alongside evaluation and optimization: it renders RunResults and optimization trajectories as markdown (render_run, render_history).

See the getting started notebook for a full working example of this cycle end to end.

Optimization

Prompt-template search (OPRO-style) treats evaluation as multi-criteria: each candidate is scored on several normalized metrics, forming a quality vector in metric space. Before comparing checkpoints, that vector is collapsed by a fixed linear scalarization—the unweighted mean of per-metric means (RunResults.overall.mean), a compensatory aggregation where gains on one metric can offset losses on another. The search is therefore single-objective in template space: it maximizes one scalar utility, keeps the best checkpoint seen so far, and does not explore a Pareto front over metrics. The proposer still receives per-metric breakdowns in the trajectory (render_history); only ranking and early stopping use the headline score.

The optimizer uses the metrics to learn the representation of the data and the expectations of the task, then encodes that knowledge in the prompt template.

Inspiration

promptuna is a proud Frankenstein of DSPy, Ragas, OPRO] and Optuna.

First and foremost, promptuna's value proposition is most similar to DSPy. The differences:

• Programs: DSPy models a program as a composable graph of predictors (dspy.Module). promptuna treats a program as an ordinary Python function with a deterministic scaffold around a single completion call, without forcing signature/module abstractions.
• Evaluation. DSPy passes a single metric callable to its optimizers. Multiple quality dimensions must be folded into that one function by hand. promptuna takes a list[Metric] instead: each metric has its own name, scale (Range, Ordinal, …), and scorer (programmatic or LLM judge). Results are naively aggregated to collapse multiple metrics into the single optimization objective.
• Optimization. DSPy offers several teleprompters. promptuna's simple optimizer is OPRO-style: it rewrites a free-form prompt template from a trajectory, using the same multi-metric evaluation harness at every step, keeping the full metric breakdown visible throughout the search.

Some ideas regarding evaluation metrics are taken from the seemingly already abandoned ragas: named metrics where an LLM judge scores a trial against a rubric, with typed scales and optional rationales.

The optimization loop itself takes concepts from DeepMind's OPRO: at each step an LM proposer rewrites the prompt template from scratch using the full scored history of prior candidates.

The name of the package itself is a reference to the infamous Optuna: a fixed-budget search over trials that archives every checkpoint and returns the best one seen.

License

MIT

Made with mold