Optimize · Azure OpenAI

Optimize Azure OpenAI

Cloud Platforms

Use Future AGI's agent-opt SDK to rewrite your Azure OpenAI prompts with measurable improvement on the metrics that matter to you.

python typescript java · trace evaluate optimize

Start free Open in docs View example on GitHub

Recipes for Azure OpenAI

Trace Azure OpenAI Evaluate Azure OpenAI Optimize Azure OpenAI live

Prerequisites

Before you start

· A working Azure OpenAI app — local or already in production.
· A free Future AGI account with FI_API_KEY and FI_SECRET_KEY.
· Python 3.9+ / Node 18+ / Java 17+ depending on which SDK you're installing.
· A dataset of ≥50 examples — Future AGI auto-builds these from your trace history.

Install

pip install traceAI-openai

npm install @traceai/openai

<dependency>
  <groupId>ai.futureagi</groupId>
  <artifactId>traceai-java-azure-openai</artifactId>
  <version>LATEST</version>
</dependency>

Optimize recipe

from agent_opt import GEPAOptimizer
from fi.evals.templates import Groundedness, PromptAdherence

optimizer = GEPAOptimizer(
    seed_prompt="<your current Azure OpenAI system prompt>",
    objectives=[Groundedness(), PromptAdherence()],
    rounds=8,
)

best_prompt, score = optimizer.run(dataset_id="azure-openai_eval_set_v1")
print(f"+{score.delta}% on grounded answers")

What Future AGI captures

Optimize fields you'll see in the dashboard

Use a Future AGI dataset of failed Azure OpenAI traces as the optimisation target
GEPA, ProTeGi, PromptWizard, MetaPrompt, Bayesian, and Random optimisers — same interface
Each optimiser run produces a new prompt version with a measured score delta
Push the best prompt back to your prompt registry and replay through the same eval suite

Common gotchas

Read these before you ship

01
Seed prompt must include the placeholder format your dataset uses (`{{question}}`, `{input}`, etc.).
02
GEPA needs ≥50 examples to converge; for smaller sets prefer ProTeGi or PromptWizard.
03
Set a hard `rounds` cap — optimisers will keep improving past your budget if you let them.

Next: chain it with the other recipes

Optimize is the first step. Most teams add an evaluator the same week, and start optimising or simulating once they have a baseline. Each recipe takes minutes to wire up.

Start free Read the Optimize docs

Adjacent integrations