optiglot 0.1.3

Make your DSPy optimizers multi-lingual.

Optiglot

optiglot is a language-agnostic interface for prompt optimization engines (like DSPy). It allows you to use powerful optimization algorithms implemented in Python with your language/framework of choice via a standardized RPC protocol. By separating the optimization logic from the execution environment, optiglot makes prompt optimizers "polyglots" hence the name.

Installation

TypeScript/JavaScript

Install the harness in your project:

npm install optiglot
# or
bun add optiglot

CLI Engine

Install the optimization engine via uv:

uv tool install optiglot

---

Core Concepts

• Harnesses: The "glue" that connects your framework (e.g., Vercel AI SDK) to the optimization engine. They handle tracing, evaluation, and communication.
• Predictors: Wrappers around LLM calls that record inputs/outputs for the optimizer.
• Module: A collection of predictors that define your AI system's logic.
• Teleprompt: The optimizer that runs the optimization loop.

---

Defining an Optimizer

In Optiglot, an optimizer is defined using the Teleprompt class. You specify which algorithm to use and its configuration.

import { Teleprompt } from "optiglot";

const optimizer = new Teleprompt("bootstrap", {
  numThreads: 4, // Number of concurrent evaluations
  maxIterations: 10, // Maximum optimization cycles
  // ... other algorithm-specific options
});

Common optimizers include:

• vanilla: Basic few-shot prompting.
• bootstrap: Bootstraps few-shot examples from your training data.
• teleprompt: Advanced multi-stage optimization.

---

The Purpose of Harnesses

Harnesses are essential because they bridge the gap between your application code and the optimization engine. They abstract away three critical tasks:

1. Tracing: Automatically recording LLM inputs and outputs during a run so the optimizer can "see" what happened.
2. Execution Control: Allowing the optimizer to swap out prompts and re-run your code during evaluation.
3. Communication: Handling the RPC handshake between your application and the Python-based optimization engine.

By using a harness, you can optimize your prompts without rewriting your existing application logic in Python.

---

Usage: AI SDK Harness

Here's how to use the Vercel AI SDK harness to optimize a support ticket classifier.

1. Define Your System (Module)

Create a Module with your predictors. A Predict component wraps your LLM call and manages its instructions.

import { Module, Predict } from "optiglot";
import { openai } from "@ai-sdk/openai";

// Define the system structure
const classifier = new Module({
  judge: new Predict(
    "Classify the ticket into: Billing, Technical, or General."
  ),
});

// Define the execution logic
classifier.setForward(async (inputs: { ticket: string }) => {
  const { text } = await classifier.judge.generateText({
    model: openai("gpt-4o-mini"),
    prompt: `Ticket: ${inputs.ticket}`,
  });
  return text;
});

2. Define Your Metric

The metric scores performance on a specific example. It can return a simple score (0 to 1) or rich feedback.

import { type MetricFunction } from "optiglot";

const metric: MetricFunction = (example, prediction) => {
  const isCorrect = example.label === prediction.output;
  return {
    score: isCorrect ? 1.0 : 0.0,
    feedback: isCorrect
      ? "Correctly labeled."
      : `Expected ${example.label} but got ${prediction.output}`,
  };
};

3. Run Optimization

Use Teleprompt to optimize the instructions in your module.

import { Teleprompt, Dataset } from "optiglot";

// Load training data
const trainset = new Dataset(
  [
    { ticket: "I can't log into my account.", label: "Technical" },
    { ticket: "Where is my order #123?", label: "Billing" },
  ],
  ["ticket"] // Fields to pass to the forward function
);

const teleprompter = new Teleprompt("bootstrap");

// Compile: This spawns the optimization engine and finds the best prompts
const optimizedModule = await teleprompter.compile(
  classifier,
  metric,
  trainset
);

console.log("Optimized Instructions:", optimizedModule.judge.instructions);

// Save the optimized state
optimizedModule.save("./optimized_prompts.json");

---

Appendix

Persistence

You can load optimized prompts back into your module for production use:

const productionModule = new Module({
  judge: new Predict(""), // Initial instructions don't matter if we're loading
});
productionModule.load("./optimized_prompts.json");

Development

If you are developing optiglot locally, you can use the OPTIGLOT_DEV=true environment variable to run the engine from source.

OPTIGLOT_DEV=true bun run your_optimization_script.ts