Reduce prompt failure rates by turning errors into executable rules
Vizpy's PromptGrad and ContraPrompt finds where your module fails, extracts the rule behind each failure, and injects precise corrections into your instructions in a single API call.
import dspy
import vizpy
# Use any model you have access to
dspy.configure(lm=dspy.LM("your_provider/model"))
# ContraPrompt — best for classification
optimizer = vizpy.ContraPromptOptimizer(metric=my_metric)
# PromptGrad — best for generation
optimizer = vizpy.PromptGradOptimizer(metric=my_metric)
# Outperforms GEPA on every benchmark
optimized = optimizer.optimize(module, train_examples)Trusted by teams at
VizPy optimizers beat GEPA across all top benchmarks.
All benchmarks are open source. Read more about how these optimizers work in our launch blogs: PromptGrad and ContraPrompt.
BBH
Naive CoT: 26.11%
HotPotQA
Naive CoT: 26.99%
GPQA Diamond
Naive CoT: 58.11%
GDPR-Bench
Naive CoT: 10.12%
Two optimizers. One clear mechanism.
Pass your module, examples, and metric. Get back a module with better instructions — and a plain-English explanation of every rule that was added.
PromptGrad
Runs epoch-based failure analysis: computes textual gradients across batches and accumulates correction rules into a separate local rules section — your base instructions stay frozen. Best for generation tasks and continuous metrics.
ContraPrompt
Mines contrastive pairs — wrong attempt vs. corrected one — and extracts the rule that separates them. Best for classification tasks with clear right/wrong labels.
Simple API
Pass any dspy.Module, get back an optimized module. Your signature and structure are unchanged — only the instructions improve.
Simple Pricing
Start free, scale as you need.