llm-extract

Extract structured data from any document using LLMs.

PDF, DOCX, HTML, CSV, plain text → validated Pydantic model, with per-field confidence scores and source grounding.

---

Why llm-extract?

Instructor extracts from strings. LangExtract extracts from text with source grounding. llm-extract does both — and handles the document loading for you.

Feature	Instructor	LangExtract	llm-extract
Pydantic schema validation	✓	✗	✓
Per-field confidence scores	✗	✗	✓
Source grounding	✗	✓	✓
Multi-format document loading	✗	✗	✓
Long document chunking	✗	✓	✓
CLI	✗	✗	✓
Provider-agnostic	✓	✗	✓ (Anthropic; more planned)

---

Install

pip install llm-extract              # core (plain text, CSV, JSON)
pip install 'llm-extract[pdf]'       # + PDF support
pip install 'llm-extract[docx]'      # + Word document support
pip install 'llm-extract[html]'      # + HTML support
pip install 'llm-extract[all]'       # everything

Set your API key:

export ANTHROPIC_API_KEY=sk-ant-...

---

Quick Start

Python API

from llm_extract import extract
from pydantic import BaseModel

class Invoice(BaseModel):
    vendor: str
    total: float
    date: str
    invoice_number: str | None = None

result = extract("invoice.pdf", schema=Invoice)

print(result.data)
# Invoice(vendor='Acme Corp', total=1250.0, date='2026-01-15', invoice_number='INV-4521')

print(result.confidence)
# {'vendor': 0.98, 'total': 0.95, 'date': 0.99, 'invoice_number': 0.92}

print(result.sources)
# {'vendor': 'Acme Corp Ltd\n123 Business Park', 'total': 'Total Due: £1,250.00', ...}

print(result.coverage)   # 1.0  (all 4 fields found)
print(result.mean_confidence)  # 0.96

From plain text

from llm_extract import extract_text
from pydantic import BaseModel

class Person(BaseModel):
    name: str
    age: int
    email: str | None = None

result = extract_text(
    "Dr. Sarah Chen, aged 34, can be reached at s.chen@nhs.uk",
    schema=Person,
)
print(result.data)
# Person(name='Dr. Sarah Chen', age=34, email='s.chen@nhs.uk')

CLI

# Extract with a schema file
llm-extract invoice.pdf --schema schemas/invoice.py

# Quick extraction — no schema file needed
llm-extract contract.txt --fields "parties,effective_date,value,governing_law"

# Show where each value was found
llm-extract report.pdf --schema schemas/report.py --show-sources

# Save result as JSON
llm-extract data.csv --fields "name,address,amount" --out result.json

# List supported formats
llm-extract formats

---

Schema Files

Define a Pydantic model in a .py file — llm-extract finds it automatically:

# schemas/contract.py
from pydantic import BaseModel, Field
from typing import Optional

class Contract(BaseModel):
    parties: list[str] = Field(description="All parties named in the contract")
    effective_date: str = Field(description="When the contract takes effect (ISO date)")
    value_gbp: Optional[float] = Field(None, description="Contract value in GBP if stated")
    governing_law: Optional[str] = Field(None, description="Governing law jurisdiction")
    termination_notice: Optional[str] = Field(None, description="Required notice period for termination")

Then:

llm-extract contract.pdf --schema schemas/contract.py --show-sources

---

Confidence Scores

Every extracted field gets a confidence score (0.0–1.0):

result = extract("document.pdf", schema=MySchema)

# Check for low-confidence fields before using the data
low = result.low_confidence_fields(threshold=0.7)
if low:
    print(f"Review these manually: {low}")

# Filter out uncertain fields
if result.confidence.get("amount", 0) < 0.8:
    print("Amount extraction uncertain — please verify")

Score	Meaning
0.95–1.0	Explicitly stated, unambiguous
0.80–0.94	Clearly implied, minor ambiguity
0.60–0.79	Inferred from context, may be wrong
Below 0.60	Very uncertain — verify manually

---

Long Documents

llm-extract automatically chunks long documents and merges results. Higher-confidence values from later chunks override lower-confidence values from earlier chunks:

from llm_extract import extract
from llm_extract.models import ExtractionConfig

config = ExtractionConfig(
    chunk_size=6000,   # characters per chunk (default: 8000)
    overlap=800,       # overlap between chunks (default: 500)
)

result = extract("long_report.pdf", schema=ReportSchema, config=config)
print(f"Processed {result.chunks_used} chunks")

---

Supported Formats

Format	Extensions	Extra Install
Plain text	.txt .md .rst	None
CSV	.csv	None
JSON	.json	None
HTML	.html .htm	pip install 'llm-extract[html]'
PDF	.pdf	pip install 'llm-extract[pdf]'
Word	.docx	pip install 'llm-extract[docx]'

---

Examples

See examples/ for complete worked examples:

• examples/invoice_extraction.py — Invoice data from PDF
• examples/contract_extraction.py — Contract terms from DOCX
• examples/clinical_notes.py — Structured data from clinical notes
• examples/csv_to_schema.py — Typed schema from CSV data

---

Comparison with LangExtract

Google's LangExtract uses few-shot examples and Gemini-specific controlled generation. llm-extract takes a different approach:

• Schema-first: define a Pydantic model, not examples
• Confidence scores: every field tells you how certain the extraction is
• Multi-format: load the document from disk, don't pre-process it yourself
• Provider-agnostic: Anthropic Claude today; OpenAI and local models planned
• CLI included: run extractions without writing Python

---

Contributing

Contributions welcome — especially:

• Additional LLM providers (OpenAI, Ollama, Gemini)
• New document format loaders
• Evaluation benchmarks on real document types

---

Linda Oraegbunam | LinkedIn | Twitter