DSPy: The framework for programming—not prompting—language models
Python
Update README.md
Omar Khattab10e6ac83ee6e
diff --git a/README.md b/README.mdindex 6e7ce4125..e5f721546 100644--- a/README.md+++ b/README.md@@ -17,9 +17,9 @@**DSPy** is a framework for developing high-quality LM systems for complex tasks. To do that without DSPy, you have to: (1) break down problems into steps, (2) prompt LMs well for each step, (3) generate synthetic data to improve quality, and (4) finetune smaller LMs to cut costs. Composing these into the right multi-step pipelines is hard. DSPy offers you the tools to automate most of this, as long as you can define your high-level steps and metrics.-**DSPy** helps by separating the flow of your program (`modules`) from the parameters (prompt instructions, few-shot examples, and LM weights), which DSPy `optimizers` can craft and tune if you give them an objective. This creates a systematic space of modular and trainable pieces, instead of hacky 'prompt engineering'. We draw on lessons from DNN frameworks like PyTorch. DSPy `modules` are layers with parameters (like `dspy.ChainOfThought`) that replace 'prompting tricks'. DSPy `optimizers` are algorithms that update parameters (like `BayesianSignatureOptimizer` or `BootstrapFewShotWithRandomSearch`) that replace prompt hacking or one-off synthetic data generators.+**DSPy** helps by separating the flow of your program (`modules`) from the parameters (prompt instructions, few-shot examples, and LM weights), which DSPy `optimizers` can craft and tune if you give them an objective. This creates a systematic space of modular and trainable pieces, instead of hacky 'prompt engineering'. We draw on lessons from DNN frameworks like PyTorch. DSPy `modules` are layers that hold parameters (like `ChainOfThought` or `ReAct`) that replace string-based prompting tricks. DSPy `optimizers` are algorithms that update parameters (like `BayesianSignatureOptimizer`) that replace prompt hacking or one-off synthetic data generators.-Using **DSPy** is an iterative process. You start by defining a task and the metrics you want to maximize. You also prepare a few example inputs, typically without labels (or only with labels for the final outputs, if your metric requires these). Then, you build your pipeline: select built-in layers (`modules`) to use, give each layer a `signature` (input/output spec), and then simply call your modules in free-form Python code to define your control flow. Lastly, use one of the DSPy `optimizers` to compile your code into high-quality instructions, automatic few-shot examples, updated LM weights for your (large or small) LM.+Using **DSPy** is an iterative process. You start by defining a task and the metrics you want to maximize. You also prepare a few example inputs, typically without labels (or only with labels for the final outputs, if your metric requires these). Then, you build your pipeline: select built-in layers (`modules`) to use, give each layer a `signature` (input/output spec), and then simply call your modules in free-form Python code to define your control flow. Lastly, use one of the DSPy `optimizers` to compile your code into high-quality instructions, automatic few-shot examples, updated LM weights for your LM.For complex tasks, **DSPy** can routinely teach powerful models like `GPT-3.5` or `GPT-4` and local models like `T5-base` or `Llama2-13b` to be much more reliable at tasks. **DSPy** will compile the _same program_ into different few-shot prompts and/or finetunes for each LM.