LLM-Powered Data Extraction and Document Processing: Patterns That Work in 2026

From Unstructured to Structured at Scale

Every enterprise sits on mountains of unstructured data: contracts, invoices, medical records, research papers, emails, support tickets. Extracting structured information from these documents has traditionally required custom NLP pipelines, regex patterns, and domain-specific models for each document type.

LLMs have changed this. A single model can extract structured data from virtually any document type with minimal customization. But doing this reliably at scale requires careful architecture.

The Basic Extraction Pattern

At its simplest, LLM-based extraction involves sending a document with a schema and asking the model to populate it:

from pydantic import BaseModel, Field
from typing import Optional

class InvoiceData(BaseModel):
    vendor_name: str
    invoice_number: str
    invoice_date: str = Field(description="ISO 8601 format")
    due_date: Optional[str] = None
    line_items: list[LineItem]
    subtotal: float
    tax: float
    total: float
    currency: str = Field(default="USD")

class LineItem(BaseModel):
    description: str
    quantity: float
    unit_price: float
    total: float

# With Anthropic's structured output
response = client.messages.create(
    model="claude-sonnet-4-20250514",
    system="Extract invoice data from the provided document. "
           "Return ONLY data explicitly stated in the document.",
    messages=[{"role": "user", "content": document_text}],
    tool_choice={"type": "tool", "name": "extract_invoice"},
    tools=[{
        "name": "extract_invoice",
        "description": "Extract structured invoice data",
        "input_schema": InvoiceData.model_json_schema()
    }]
)

Chunking Strategies for Long Documents

Documents that exceed the model's context window (or are too expensive to process whole) need chunking. But naive chunking breaks extraction because relevant information may span chunk boundaries.

Sliding Window with Overlap:

flowchart TD
    HUB(("From Unstructured to<br/>Structured at Scale"))
    HUB --> L0["The Basic Extraction Pattern"]
    style L0 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L1["Chunking Strategies for Long<br/>Documents"]
    style L1 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L2["Handling Multi-Page<br/>Documents"]
    style L2 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L3["Quality Assurance Patterns"]
    style L3 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L4["Production Architecture"]
    style L4 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L5["Cost Optimization"]
    style L5 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    style HUB fill:#4f46e5,stroke:#4338ca,color:#fff

def chunk_document(text, chunk_size=3000, overlap=500):
    chunks = []
    start = 0
    while start < len(text):
        end = start + chunk_size
        chunks.append(text[start:end])
        start = end - overlap
    return chunks

Section-Aware Chunking: Parse the document structure first (headings, tables, paragraphs) and chunk at logical boundaries. This preserves the semantic integrity of each chunk.

Two-Pass Extraction: First pass identifies which sections contain relevant information. Second pass extracts from only those sections.

Handling Multi-Page Documents

For complex documents like contracts or medical records:

1. Page-level extraction: Extract data from each page independently
2. Merge and deduplicate: Combine results across pages, resolving conflicts
3. Cross-reference validation: Check extracted values for consistency (e.g., does the sum of line items equal the total?)

async def extract_from_document(pages: list[str], schema: type[BaseModel]):
    # Extract from each page in parallel
    page_results = await asyncio.gather(*[
        extract_page(page, schema) for page in pages
    ])

    # Merge results with conflict resolution
    merged = merge_extractions(page_results, strategy="highest_confidence")

    # Validate consistency
    validation_errors = validate_extraction(merged)
    if validation_errors:
        # Re-extract with targeted prompts for inconsistent fields
        merged = await resolve_conflicts(merged, validation_errors, pages)

    return merged

Quality Assurance Patterns

Confidence Scoring

Ask the model to rate its confidence for each extracted field:

class ExtractedField(BaseModel):
    value: str
    confidence: float = Field(ge=0, le=1, description="Extraction confidence 0-1")
    source_text: str = Field(description="Exact text from document supporting this value")

Route low-confidence extractions to human review.

Dual Extraction

Run extraction twice (potentially with different models or prompts) and compare results. Disagreements flag potential errors:

• Both agree: high confidence, auto-accept
• One extraction has the field, other does not: medium confidence, review if critical
• Both have different values: low confidence, always route to human review

Schema Validation

Use Pydantic validators to catch impossible values:

from pydantic import validator

class InvoiceData(BaseModel):
    total: float
    line_items: list[LineItem]

    @validator('total')
    def total_matches_line_items(cls, v, values):
        if 'line_items' in values:
            expected = sum(item.total for item in values['line_items'])
            if abs(v - expected) > 0.01:
                raise ValueError(f"Total {v} doesn't match sum of line items {expected}")
        return v

Production Architecture

A production document processing pipeline typically looks like:

Document Upload -> OCR (if scanned) -> Text Extraction
    -> Classification (what type of document?)
    -> Schema Selection (which extraction schema to use?)
    -> Chunking -> Parallel Extraction -> Merge -> Validation
    -> Confidence Routing:
        High confidence -> Auto-accept -> Database
        Low confidence -> Human Review Queue -> Database

Cost Optimization

Document extraction can be expensive at scale. Optimize by:

• Using cheaper models (Haiku, GPT-4o mini) for classification and simple extractions
• Reserving expensive models for complex documents or low-confidence re-extraction
• Caching extraction results for identical documents (hash-based dedup)
• Batch processing during off-peak hours for non-urgent documents

Sources: Anthropic Structured Output | LlamaIndex Document Processing | Unstructured.io

flowchart LR
    IN(["Input prompt"])
    subgraph PRE["Pre processing"]
        TOK["Tokenize"]
        EMB["Embed"]
    end
    subgraph CORE["Model Core"]
        ATTN["Self attention layers"]
        MLP["Feed forward layers"]
    end
    subgraph POST["Post processing"]
        SAMP["Sampling"]
        DETOK["Detokenize"]
    end
    OUT(["Generated text"])
    IN --> TOK --> EMB --> ATTN --> MLP --> SAMP --> DETOK --> OUT
    style IN fill:#f1f5f9,stroke:#64748b,color:#0f172a
    style CORE fill:#ede9fe,stroke:#7c3aed,color:#1e1b4b
    style OUT fill:#059669,stroke:#047857,color:#fff

flowchart TD
    HUB(("From Unstructured to<br/>Structured at Scale"))
    HUB --> L0["The Basic Extraction Pattern"]
    style L0 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L1["Chunking Strategies for Long<br/>Documents"]
    style L1 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L2["Handling Multi-Page<br/>Documents"]
    style L2 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L3["Quality Assurance Patterns"]
    style L3 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L4["Production Architecture"]
    style L4 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    HUB --> L5["Cost Optimization"]
    style L5 fill:#e0e7ff,stroke:#6366f1,color:#1e293b
    style HUB fill:#4f46e5,stroke:#4338ca,color:#fff