SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsdspy-predict

dspy-predict

General & other

View repository

Updated 10-15 23:58

About this skill

Direct LM Calls with dspy.Predict

Guide the user through using dspy.Predict -- the simplest and fastest DSPy module for calling a language model. It takes inputs, calls the LM, and returns typed outputs. No intermediate reasoning, no extra steps.

Step 1: Gather context

Before writing code, clarify:

  1. What is the task? Classification, extraction, formatting, or Q&A?
  2. What output type is needed? A plain string, a constrained Literal list, or a structured Pydantic model?
  3. Do you need batch processing? Thousands of items at once, or single calls?
  4. Is there an existing signature, or does it need to be designed from scratch? (If from scratch, see /dspy-signatures.)

What is dspy.Predict

dspy.Predict is the atomic building block of every DSPy program -- one LM call, no reasoning chain, no tool loops. It takes a signature and calls the LM once to produce the output fields. Every other DSPy module (ChainOfThought, ReAct, etc.) builds on top of it.

When to use Predict vs ChainOfThought

Use dspy.Predict when... Use dspy.ChainOfThought when...
The task is straightforward (classification, extraction, formatting) The task benefits from step-by-step reasoning
You want minimal latency and token usage Accuracy matters more than speed
The mapping from input to output is direct The LM needs to "think through" intermediate steps
You're building a simple sub-step inside a larger pipeline You need to inspect the model's reasoning

Rule of thumb: Start with Predict. If accuracy is too low, switch to ChainOfThought -- it's a one-word change.

Predict with Pydantic output types

For complex structured outputs, use a Pydantic BaseModel as the output type:

import dspy
from pydantic import BaseModel

lm = dspy.LM("openai/gpt-4o-mini")  # or "anthropic/claude-sonnet-4-5-20250929", etc.
dspy.configure(lm=lm)

class ContactInfo(BaseModel):
    name: str
    email: str
    phone: str

class ExtractContact(dspy.Signature):
    """Extract contact information from the text."""
    text: str = dspy.InputField()
    contact: ContactInfo = dspy.OutputField()

extractor = dspy.Predict(ExtractContact)
result = extractor(text="Reach out to Jane Doe at jane@example.com or 555-0123")
print(result.contact.name)   # Jane Doe
print(result.contact.email)  # jane@example.com
print(result.contact.phone)  # 555-0123

Async and batch processing

For async calls, use acall or aforward:

result = await predict.acall(question="What is DSPy?")

For batch processing, use the built-in batch() method instead of a Python loop:

examples = [dspy.Example(question=q).with_inputs("question") for q in questions]
results = predict.batch(examples, num_threads=8, timeout=120)

Save and load optimized predictors:

predict.save("my_predictor.json")
predict.load("my_predictor.json")

Verify your predictor

After building, spot-check on a small eval set before scaling up:

# Quick sanity check (a few labeled examples)
for ex in devset[:10]:
    result = predict(**ex.inputs())
    status = "PASS" if result.label == ex.label else "FAIL"
    print(f"{status}: expected={ex.label!r} got={result.label!r}")

# Larger eval with dspy.Evaluate
from dspy.evaluate import Evaluate
evaluator = Evaluate(devset=devset, metric=your_metric, num_threads=4)
score = evaluator(predict)
print(f"Accuracy: {score:.1%}")

If accuracy is below target, optimize before switching modules — see /dspy-watching-optimization for guidance.

Gotchas

  1. Switching to ChainOfThought as the first fix when accuracy is low. When Predict underperforms, the instinct is to swap modules. But a Predict compiled with BootstrapFewShot or MIPROv2 typically matches or beats bare ChainOfThought at lower token cost. Optimize Predict first; switch to ChainOfThought only when the task genuinely requires intermediate reasoning steps.
  2. Forgetting .with_inputs() on examples passed to batch(). Every dspy.Example must call .with_inputs("field_name") to mark which fields are inputs vs. outputs. Omitting it causes DSPy to treat all fields as outputs, breaking batch execution and optimizer training.
  3. For batch processing, use predict.batch() rather than a Python loop -- it uses dspy.Parallel internally, handles concurrency, and is significantly faster for large batches.
  4. Only keyword arguments -- predict("my input") raises ValueError. Always use predict(question="my input").

Cross-references

Install any skill: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill <name>

  • Defining signatures (inline and class-based, typed fields, Pydantic outputs) -- see /dspy-signatures
  • Adding step-by-step reasoning -- see /dspy-chain-of-thought
  • Building multi-step programs with modules that compose Predict calls -- see /dspy-modules
  • Classification and sorting with real-world patterns -- see /ai-sorting
  • For worked examples, see examples.md
  • Install /ai-do if you do not have it — it routes any AI problem to the right skill and is the fastest way to work: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill ai-do

Additional resources