dspy

DSPy Expert Skill

DSPy is a compiler for prompt programs, not a chain or RAG framework. You write Python programs with typed signatures and DSPy optimizes the prompts automatically.

⚠️ DSPy is NOT a chain framework. It does not use prompt | model | parser. It does not have LCEL. DSPy operates at a different layer: you define a program with Python control flow and typed signatures, then the compiler optimizes the prompts against a metric. If you reach for DSPy expecting LangChain-style composition, you are reaching for the wrong tool.

Think of it as PyTorch for LMs — you define the architecture, the compiler tunes the weights (prompts).

Core Paradigm

Read this first. It is the most important thing to understand about DSPy.

import dspy

# 1. Configure the LM
lm = dspy.LM("openai/gpt-4o-mini")
dspy.configure(lm=lm)

# 2. Define a signature (input/output schema)
class QASignature(dspy.Signature):
    """Answer questions concisely."""
    question: str = dspy.InputField()
    answer: str = dspy.OutputField()

# 3. Build a program using modules
qa = dspy.ChainOfThought(QASignature)

# 4. Compile against a metric
optimizer = dspy.MIPROv2(metric=dspy.answer_exact_match)
compiled_qa = optimizer.compile(qa, trainset=trainset, num_trials=25)

# 5. Use the compiled program (portable artifact)
answer = compiled_qa(question="What is DSPy?").answer

Core Principles

1. DSPy is a compiler, not a chain framework. You define the program structure with Python control flow and typed signatures. The compiler optimizes the prompts. This is fundamentally different from LangChain's explicit prompt composition.

2. Signatures define the task. Input/output field pairs with optional descriptions are the task definition. The syntax is input1, input2 -> output1, output2.

3. Modules are program components. dspy.Predict (direct), dspy.ChainOfThought (reasoning), dspy.ReAct (tool-use), and custom dspy.Module subclasses. Compose them with Python control flow (if/for/while).