TextGrad

// Overview

TextGrad is a Python framework that brings automatic "differentiation" to text. Instead of numeric gradients, it backpropagates feedback written by an LLM, so you can improve prompts, answers, code, and other text variables through a loop that mirrors how neural networks are trained.

The API deliberately follows PyTorch. You wrap text in a Variable, define a loss with TextLoss, and call loss.backward() and optimizer.step() using TGD (Textual Gradient Descent). If you know PyTorch, most of the concepts carry over directly.

It fits the prompt-programming space as an optimization layer: rather than hand-tuning prompts, you let the framework critique and revise text against an evaluation instruction. It works with many model providers through a litellm-based engine, including OpenAI, Anthropic, Gemini, Bedrock, and Together.

// What it does

• PyTorch-style API: Variable, TextLoss, backward(), and a TGD (Textual Gradient Descent) optimizer
• Backpropagation through natural-language feedback produced by an LLM
• Optimizes many kinds of text variables — prompts, answers, code, and solutions
• BlackboxLLM wrapper for running a model forward pass on a Variable
• Experimental litellm engine supporting OpenAI, Anthropic, Gemini, Bedrock, Together, and more
• Optional response caching and multimodal (image) input via the litellm engine

// Getting started

Install the package, set your model API key, then run a short optimize-the-answer loop that mirrors PyTorch.

pip install textgrad

import textgrad as tg

tg.set_backward_engine("gpt-4o", override=True)

model = tg.BlackboxLLM("gpt-4o")
question_string = ("If it takes 1 hour to dry 25 shirts under the sun, "
                   "how long will it take to dry 30 shirts under the sun? "
                   "Reason step by step")

question = tg.Variable(question_string,
                       role_description="question to the LLM",
                       requires_grad=False)

answer = model(question)

answer.set_role_description("concise and accurate answer to the question")

optimizer = tg.TGD(parameters=[answer])
evaluation_instruction = (f"Here's a question: {question_string}. "
                          "Evaluate any given answer to this question, "
                          "be smart, logical, and very critical. "
                          "Just provide concise feedback.")

loss_fn = tg.TextLoss(evaluation_instruction)

loss = loss_fn(answer)
loss.backward()
optimizer.step()

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Automatically refine an LLM's reasoning or answer when a first response is wrong
• Tune system prompts and instructions instead of editing them by hand
• Improve generated code or problem solutions through iterative LLM critique
• Optimize text variables in research workflows across multiple model providers via litellm

// How TextGrad compares

TextGrad alongside other open-source prompt programming tools AI/TLDR tracks, ranked by GitHub stars.