ai-parsing-data

name

ai-parsing-data

description

Pull structured data from messy text using AI. Use when parsing invoices, extracting fields from emails, scraping entities from articles, converting unstructured text to JSON, extracting contact info, parsing resumes, reading forms, pulling data from transcripts (VTT, LiveKit, Recall), extracting fields from Langfuse traces, or any task where messy text goes in and clean structured data comes out. Also use when emails are messy and lack structure, or structured data extraction from unstructured content is unreliable., extract entities from text, parse PDF with AI, structured extraction from unstructured text, OCR plus AI extraction, convert email to structured data, pull fields from documents automatically, AI data entry automation, invoice parsing, resume parsing with AI, medical record extraction.

Build an AI Data Parser

Guide the user through building AI that pulls structured data out of messy text. Uses DSPy extraction — define the output shape, and the AI fills it in.

Step 1: Define what to extract

Ask the user:

1. What are you parsing? (emails, invoices, resumes, transcripts, articles, forms, etc.)
2. What fields do you need? (names, dates, amounts, entities, etc.)
3. Are any fields optional? (some documents might not have every field)
4. What's the output format? (flat fields, list of objects, nested structure)
5. Do you have examples of correct extractions? (even a few help with optimization)

Step 2: Build the parser

Simple field extraction

For pulling a known set of fields from text:

import dspy

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

class ParseContact(dspy.Signature):
    """Extract contact information from the text."""
    text: str = dspy.InputField(desc="Text containing contact information")
    name: str = dspy.OutputField(desc="Person's full name")
    email: str = dspy.OutputField(desc="Email address")
    phone: str = dspy.OutputField(desc="Phone number")

parser = dspy.ChainOfThought(ParseContact)

ChainOfThought adds reasoning before extraction, which helps the model think through which text maps to which field — typically 5-15% more accurate than bare Predict on ambiguous inputs.

Structured output with Pydantic

For complex or nested output, use Pydantic models. DSPy handles the serialization automatically:

from pydantic import BaseModel, Field
from typing import Optional

class Address(BaseModel):
    street: str
    city: str
    state: str
    zip_code: str

class Person(BaseModel):
    name: str
    age: Optional[int] = None
    email: Optional[str] = None
    address: Address
    skills: list[str]

class ParsePerson(dspy.Signature):
    """Extract person details from the text."""
    text: str = dspy.InputField()
    person: Person = dspy.OutputField()

parser = dspy.ChainOfThought(ParsePerson)
result = parser(text="John Doe, 32, lives at 123 Main St, Springfield IL 62701. Expert in Python and SQL.")
print(result.person)  # Person(name='John Doe', age=32, ...)

Use Optional for fields that might not appear in every document — this tells the model it's OK to return None instead of guessing.

Output format for small models

Small models (<4B params) produce frequent JSON syntax errors — unclosed braces, missing quotes, trailing commas. Switching to YAML output eliminates these failures entirely while preserving structured data. In one production case (3.6M historical name records), frontier models achieved ~70% accuracy while fine-tuned 0.8B-4B models using YAML output hit 94-96%.

import yaml

class ParsePersonYAML(dspy.Signature):
    """Extract person details from the text. Return the result as YAML, not JSON."""
    text: str = dspy.InputField()
    person_yaml: str = dspy.OutputField(desc="extracted person data in YAML format")

parser = dspy.Predict(ParsePersonYAML)
result = parser(text="John Doe, 32, [e-mail]")

# Parse YAML back into structured data
person_data = yaml.safe_load(result.person_yaml)

Use this pattern when running sub-4B parameter models (Qwen, Phi, Gemma) locally. Larger models (GPT-4o, Claude) handle JSON fine — stick with Pydantic output fields for those.

List extraction

When you need to pull a variable number of items (entities, line items, experiences):

class Entity(BaseModel):
    name: str
    type: str = Field(description="Type: person, organization, location, or date")

class ParseEntities(dspy.Signature):
    """Extract all named entities from the text."""
    text: str = dspy.InputField()
    entities: list[Entity] = dspy.OutputField(desc="All entities found in the text")

parser = dspy.ChainOfThought(ParseEntities)

Step 3: Load your data

From files

from pathlib import Path

# Single file
text = Path("document.txt").read_text()
result = parser(text=text)

# Directory of files
documents = []
for path in Path("documents/").glob("*.txt"):
    documents.append({"file": path.name, "text": path.read_text()})

From a CSV

import pandas as pd

df = pd.read_csv("emails.csv")  # column: body
results = []
for _, row in df.iterrows():
    result = parser(text=row["body"])
    results.append(result.person.model_dump())  # Pydantic → dict

# Save extracted data
pd.DataFrame(results).to_csv("extracted.csv", index=False)

From transcripts (VTT, LiveKit, Recall)

Transcripts are a common parsing source — extracting caller info, action items, decisions, or structured summaries from conversations.

WebVTT (.vtt) files:

import re

def load_vtt(path):
    """Extract text from a VTT transcript, stripping timestamps."""
    text = open(path).read()
    lines = [line.strip() for line in text.split("\n")
             if line.strip() and not line.startswith("WEBVTT")
             and not re.match(r"\d{2}:\d{2}", line)
             and not line.strip().isdigit()]
    return " ".join(lines)

LiveKit transcripts:

import json

def load_livekit_transcript(path):
    """Extract text from a LiveKit transcript JSON export."""
    data = json.load(open(path))
    segments = data.get("segments", data.get("results", []))
    return " ".join(seg.get("text", "") for seg in segments)

Recall.ai transcripts:

def load_recall_transcript(transcript_data):
    """Extract text from a Recall.ai transcript response."""
    return " ".join(
        entry["words"] for entry in transcript_data if entry.get("words")
    )

Example: extracting structured data from a call transcript:

class CallSummary(BaseModel):
    caller_name: Optional[str] = None
    issue_summary: str
    resolution: Optional[str] = None
    follow_up_needed: bool
    action_items: list[str]

class ParseCallTranscript(dspy.Signature):
    """Extract structured information from a customer call transcript."""
    transcript: str = dspy.InputField(desc="Full call transcript text")
    summary: CallSummary = dspy.OutputField()

parser = dspy.ChainOfThought(ParseCallTranscript)
transcript = load_livekit_transcript("call_001.json")
result = parser(transcript=transcript)

From Langfuse traces

Extract structured data from AI interactions logged in Langfuse:

from langfuse import Langfuse

langfuse = Langfuse()
traces = langfuse.fetch_traces(limit=100).data

# Parse each trace's input/output for structured fields
for trace in traces:
    if trace.input:
        text = trace.input.get("message", str(trace.input))
        result = parser(text=text)

Step 4: Handle messy data

Real-world text is messy. Use a reward function with dspy.Refine to catch bad extractions and retry:

class ValidatedParser(dspy.Module):
    def __init__(self):
        self.parse = dspy.ChainOfThought(ParseContact)

    def forward(self, text):
        return self.parse(text=text)

def contact_reward(args, pred):
    score = 1.0
    if not pred.email or "@" not in pred.email:
        score -= 0.4  # Email should contain @
    phone_digits = pred.phone.replace("-", "").replace(" ", "") if pred.phone else ""
    if len(phone_digits) < 10:
        score -= 0.3  # Phone number should have at least 10 digits
    return max(score, 0.0)

validated_parser = dspy.Refine(
    module=ValidatedParser(),
    N=3,
    reward_fn=contact_reward,
    threshold=0.7,
)

dspy.Refine retries the extraction up to N times, keeping the attempt with the highest reward score. Penalize each failed constraint proportionally to its importance.

Hybrid extraction with regex backstop

The model does the heavy lifting, then regex sweeps for anything it missed. This pattern improved F1 from 0.733 (model-only) to 0.929 (hybrid) in a production privacy extraction system.

import re

def hybrid_extract(text, parser, patterns):
    """Run model extraction, then fill gaps with regex patterns."""
    result = parser(text=text)

    # Regex backstop — catch fields the model missed
    for field, pattern in patterns.items():
        model_value = getattr(result, field, None)
        if not model_value:
            match = re.search(pattern, text)
            if match:
                result.__dict__[field] = match.group(1) if match.groups() else match.group()

    return result

# Define regex patterns for common fields
patterns = {
    "email": r"[\w.+-]+@[\w-]+\.[\w.-]+",
    "phone": r"\(?\d{3}\)?[-.\s]?\d{3}[-.\s]?\d{4}",
    "zip_code": r"\b\d{5}(?:-\d{4})?\b",
}

result = hybrid_extract(messy_text, parser, patterns)

Use this when your input has a mix of structured patterns (emails, phones, dates) and free-form text. The model handles ambiguous fields like names and summaries; regex catches well-formatted fields the model occasionally skips.

Handling missing fields

When a field genuinely isn't in the text, you want the model to say so rather than hallucinate a value. Use Optional types in your Pydantic model, and add a validation note in the signature docstring:

class ParseContact(dspy.Signature):
    """Extract contact info from the text. Return None for fields not present — do not guess."""
    text: str = dspy.InputField()
    name: str = dspy.OutputField(desc="Person's full name")
    email: Optional[str] = dspy.OutputField(desc="Email address, or None if not found")
    phone: Optional[str] = dspy.OutputField(desc="Phone number, or None if not found")

Step 5: Evaluate quality

from dspy.evaluate import Evaluate

def parsing_metric(example, prediction, trace=None):
    """Score based on field-level accuracy (partial credit)."""
    correct = 0
    total = 0
    for field in ["name", "email", "phone"]:
        expected = getattr(example, field, None)
        predicted = getattr(prediction, field, None)
        if expected is not None:
            total += 1
            if predicted and expected.lower().strip() == predicted.lower().strip():
                correct += 1
    return correct / total if total > 0 else 0.0

evaluator = Evaluate(devset=devset, metric=parsing_metric, num_threads=4, display_progress=True)
score = evaluator(parser)
print(f"Baseline accuracy: {score}%")

For Pydantic outputs, compare field-by-field or use the model's .model_dump() to compare dicts. Partial credit (scoring each field independently) is better than all-or-nothing for extraction tasks — it tells you which specific fields are causing problems.

Step 6: Optimize and deploy

# Optimize
optimizer = dspy.BootstrapFewShot(metric=parsing_metric, max_bootstrapped_demos=4)
optimized = optimizer.compile(parser, trainset=trainset)

# Evaluate improvement
improved = evaluator(optimized)
print(f"Optimized accuracy: {improved}%")

# Save for production
optimized.save("parser.json")

# Load later
parser = dspy.ChainOfThought(ParseContact)
parser.load("parser.json")

Batch processing

For parsing many documents at once:

import json

results = []
errors = []

for doc in documents:
    try:
        result = optimized(text=doc["text"])
        results.append({
            "source": doc["file"],
            **result.person.model_dump()  # flatten Pydantic fields
        })
    except Exception as e:
        errors.append({"source": doc["file"], "error": str(e)})

# Save results
with open("extracted.json", "w") as f:
    json.dump(results, f, indent=2)

if errors:
    print(f"{len(errors)} documents failed to parse — check errors list")

Additional resources

• For worked examples (invoices, resumes, entities, relations, forms), see examples.md
• Need summaries instead of structured data? Use /ai-summarizing
• AI missing items on complex inputs? Use /ai-decomposing-tasks
• Want to measure and improve further? Use /ai-improving-accuracy
• Need to generate training data? Use /ai-generating-data

Gotchas

• Pydantic models must be JSON-serializable — avoid custom types, datetime objects, or complex validators in output models. Stick to str, int, float, bool, list, dict, and nested Pydantic models.
• Optional fields need an explicit "return None" instruction — use field: Optional[str] = dspy.OutputField(desc="... or None if not found") and state "return None for missing fields" in the signature docstring, or the model will hallucinate values for missing fields instead of returning None.
• List extraction undercounts by default — when extracting lists of items (e.g., "all people mentioned"), the LM tends to stop early. Set max_tokens higher and add a "be exhaustive" instruction in the signature docstring.
• Long inputs get truncated silently — if your input text exceeds the model's context window, DSPy doesn't warn you. Chunk long documents before parsing, or use a model with a larger context window.
• Nested Pydantic models increase failure rate — each level of nesting adds extraction difficulty. Flatten where possible, or break into multiple extraction steps (extract outer structure first, then fill in nested fields).
• 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