Extract Free-form Text

Learn how to extract unstructured text output without defining a Pydantic model.

When You'd Use This

Not every extraction task needs a rigid schema. Sometimes you want:

• Sentiment analysis — just a sentiment label or score, not a structured object
• Text summarization — a free-form summary of a document
• Open-ended Q&A — answers that vary in structure
• Classification — assigning one of several text categories

Instead of a Pydantic model with typed fields, you optimize a single text output using model=None.

Step 1: Create Examples

Prepare examples with string expected_output:

from dspydantic import Example

examples = [
    Example(
        text="The movie was absolutely brilliant. Amazing performances and a gripping story.",
        expected_output="positive"
    ),
    Example(
        text="Terrible acting, boring plot, waste of time.",
        expected_output="negative"
    ),
    Example(
        text="It was okay. Some good parts, some slow scenes.",
        expected_output="neutral"
    ),
    Example(
        text="A masterpiece! Every scene was perfect. Highly recommend it.",
        expected_output="positive"
    ),
    Example(
        text="Not what I expected. Didn't really enjoy it.",
        expected_output="negative"
    ),
]

Tips: - Every example must have expected_output as a string, not a dict - Use multiple examples (5-20 recommended) to teach the model your desired output format - Examples should be representative of real data you'll process

Step 2: Configure and Optimize

import dspy
from dspydantic import Prompter

# Configure language model (see Configure a Language Model tutorial)
dspy.configure(lm=dspy.LM("openai/gpt-4o-mini", api_key="your-api-key"))

# Create prompter with model=None
prompter = Prompter(model=None)

# Optimize with your examples
result = prompter.optimize(examples=examples)

print(f"Baseline accuracy: {result.baseline_score:.0%}")
print(f"Optimized accuracy: {result.optimized_score:.0%}")

Output example:

Baseline accuracy: 60%
Optimized accuracy: 92%

The optimizer improved the field description and system/instruction prompts to better guide the model toward your desired output format.

Step 3: Extract

Use your optimized prompter to extract from new text:

data = prompter.run("""
    I loved this film. The director did an amazing job,
    and the cinematography was stunning.
""")

print(data.output)  # "positive"

Access the result with .output — that's the string you optimized for.

With Images and PDFs

The same pattern works with images and PDFs. Just use the appropriate input format:

Images

examples = [
    Example(image_path="digit_5.png", expected_output="5"),
    Example(image_path="digit_3.png", expected_output="3"),
    Example(image_path="digit_7.png", expected_output="7"),
]

prompter = Prompter(model=None)
result = prompter.optimize(examples=examples)

digit = prompter.run(image_path="new_digit.png")
print(digit.output)  # e.g., "5"

PDFs

examples = [
    Example(pdf_path="invoice_001.pdf", expected_output="INV-2024-001"),
    Example(pdf_path="invoice_002.pdf", expected_output="INV-2024-002"),
]

prompter = Prompter(model=None)
result = prompter.optimize(examples=examples)

invoice = prompter.run(pdf_path="new_invoice.pdf")
print(invoice.output)  # e.g., "INV-2024-003"

How It Works

Under the hood, DSPydantic creates a minimal internal schema with a single field "output" (a string field). When you optimize, it:

1. Optimizes the field description for "output" to clarify what you want
2. Optimizes the system and instruction prompts
3. Tests against your examples to measure accuracy

When you call prompter.run(...), the model generates text and DSPydantic extracts the string result.

What Gets Optimized

What	Impact
"output" field description	High — describes the desired output format
System/instruction prompts	Medium — guides the overall extraction behavior

Tips

• Use consistent output formats across examples (e.g., always "positive", "negative", "neutral" for sentiment)
• When output should be numeric or structured, consider using a Pydantic model instead (see Extract Structured Data)
• For complex multi-step reasoning, use Prompt Templates for dynamic prompts
• To save your optimized prompter for production, see Save and Load a Prompter

Next Steps

Topic	Guide
Structured data with types	Extract Structured Data
Dynamic prompts with placeholders	Optimize with Prompt Templates
Images and PDFs in detail	Use Images and PDFs
Production deployment	Save and Load a Prompter
Customize evaluation	Configure Evaluators