Skip to main content
MIPRO optimizes prompts through systematic instruction proposal and evaluation.

​
Quick Start (Online Mode)

MIPRO supports online optimization where you drive rollouts locally and the backend provides prompt candidates: 
[prompt_learning]
algorithm = "mipro"
task_app_url = "https://your-tunnel.trycloudflare.com"

[prompt_learning.mipro]
mode = "online"
bootstrap_train_seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
val_seeds = [100, 101, 102, 103, 104]
online_pool = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
online_proposer_mode = "inline"
online_proposer_min_rollouts = 20
online_rollouts_per_candidate = 10

[prompt_learning.mipro.proposer]
mode = "instruction_only"
model = "gpt-4o-mini"
provider = "openai"
temperature = 0.7
max_tokens = 512

[prompt_learning.policy]
model = "gpt-4o-mini"
provider = "openai"
temperature = 0.0
max_completion_tokens = 256
Run with: 
import os
from synth_ai.sdk import PolicyOptimizationJob

def run_mipro():
    job = PolicyOptimizationJob.from_config(
        "my_config.toml",
        api_key=os.environ["SYNTH_API_KEY"],
        algorithm="mipro"
    )
    job.submit()
    result = job.poll_until_complete()
    print(f"Best score: {result.best_score}")

​
Online vs Offline Mode

MIPRO supports two optimization modes:

​
Online Mode

  • You control rollouts: Drive the rollout loop locally
  • No tunneling required: Backend never calls your task app
  • Real-time evolution: Prompts evolve as rewards are reported
  • Proxy URL: Backend provides a proxy URL that selects prompt candidates for each LLM call

​
Offline Mode

  • Backend drives rollouts: Backend calls your task app
  • Tunneling required: Task app must be publicly accessible
  • Environment API key: ENVIRONMENT_API_KEY must be registered with backend

​
Required Fields

FieldDescription
task_app_urlURL of your task app (for offline mode) or any URL (for online mode)
mipro.modeOptimization mode: "online" or "offline"
mipro.bootstrap_train_seedsInitial training seeds for bootstrap phase
mipro.val_seedsValidation seeds for evaluation
mipro.online_poolPool of seeds for online optimization (online mode only)
mipro.online_proposer_modeProposer mode: "inline" (online mode only)
mipro.online_proposer_min_rolloutsMinimum rollouts before generating new proposals (online mode only)
mipro.online_rollouts_per_candidateNumber of rollouts per candidate before switching (online mode only)
mipro.proposer.modeProposer generation mode: "instruction_only"
mipro.proposer.modelModel for generating proposals
mipro.proposer.providerProvider for proposer model
policy.modelModel for policy execution
policy.providerProvider for policy model

​
Proposer Configuration

The proposer generates new prompt instructions. Key parameters:
  • proposer.mode: Currently supports "instruction_only" for instruction-level proposals
  • proposer.model: Model used for generating proposals (typically gpt-4o-mini or gpt-4o)
  • proposer.temperature: Controls randomness (0.7 recommended)
  • proposer.max_tokens: Maximum tokens for proposal output (512 recommended)
The proposer API key is automatically resolved from the backend’s environment (OPENAI_API_KEY or PROD_OPENAI_API_KEY).

​
Online Mode Workflow

  1. Create job: Submit MIPRO job with mode: "online"
  2. Get proxy URL: Backend returns a proxy URL endpoint
  3. Run rollouts: For each rollout:
    • Call proxy URL with your task input
    • Proxy selects best prompt candidate
    • Execute LLM call with selected prompt
    • Report reward back to backend
  4. Automatic evolution: Backend generates new proposals based on rewards

​
See Also