dspy-2-4-12

name	dspy-2-4-12
description	Framework for programming rather than prompting language models. Compiles LM calls into self-improving pipelines by tuning prompts and/or weights to maximize user-defined metrics. Use when building classifiers, RAG pipelines, agents, or any multi-stage LM program requiring automated prompt engineering, few-shot bootstrapping, or model fine-tuning driven by evaluation metrics.
license	MIT
author	Tangled <[e-mail]>
version	0.2.0
tags	["llm-programming","prompt-optimization","ai-compilation","few-shot-learning","instruction-tuning","agent-framework","dspy"]
category	ai-framework
external_references	["https://github.com/stanfordnlp/dspy/tree/v2.4.12"]

DSPy 2.4.12

Overview

DSPy (Declarative Self-improving Python) is the framework for programming—rather than prompting—language models. Instead of writing brittle, hand-crafted prompts, you write compositional Python code using declarative modules and let DSPy's optimizers teach your language model to deliver high-quality outputs.

DSPy provides three core abstractions:

• Signatures — Declarative specifications of input/output behavior that tell the LM what to do without specifying how.
• Modules — Building blocks that abstract prompting techniques (chain-of-thought, ReAct, program-of-thought) and can be composed into larger programs.
• Optimizers (formerly Teleprompters) — Algorithms that tune prompts and/or LM weights to maximize user-defined metrics like accuracy.

When to Use

• Building classifiers that need automated prompt optimization instead of manual tuning
• Creating RAG pipelines where retrieval and generation steps benefit from compiled prompts
• Implementing agent loops (ReAct, ProgramOfThought) with tools and self-correction
• Any multi-stage LM program where you want to optimize few-shot examples, instructions, or model weights driven by evaluation metrics
• Fine-tuning small LMs on task-specific data using DSPy's BootstrapFinetune optimizer
• Enforcing computational constraints on LM outputs with DSPy Assertions
• Building typed LM programs with TypedPredictor for Pydantic-compatible structured outputs

Installation

pip install dspy-ai

Optional extras:

pip install dspy-ai[chromadb]   # ChromaDB
pip install dspy-ai[qdrant]     # Qdrant
pip install dspy-ai[milvus]     # Milvus
pip install dspy-ai[pinecone]   # Pinecone
pip install dspy-ai[weaviate]   # Weaviate
pip install dspy-ai[faiss-cpu]  # FAISS
pip install dspy-ai[groq]       # Groq

Core Concepts

Signatures

A Signature is a declarative specification of input/output behavior.

Inline signatures use short string notation: