Parallel Extraction

Parallel Extraction is an advanced extraction mode that dramatically speeds up processing for documents containing repeating data structures, such as invoices with multiple line items, tables with many rows, or any document with arrays of similar objects.

How It Works

Instead of processing an entire document in a single LLM call, Parallel Extraction:

1. Identifies array structures in your JSON schema (e.g., line_items, transactions, entries)
2. Extracts unique keys from the document (e.g., product names, invoice numbers, dates) that identify each item
3. Segments the document by locating where each key appears
4. Processes items concurrently - each array item is extracted in parallel using dedicated LLM calls
5. Merges results into the final structured output while preserving order

This approach is particularly effective for multi-page documents where items span across pages, as each segment can be processed independently.

Context Engineering: Why Parallel Extraction is More Accurate

Beyond speed, Parallel Extraction significantly improves extraction accuracy through Context Engineering — the practice of optimizing what context the LLM sees for each extraction task.

When processing a 50-page invoice with 200 line items in a single LLM call, the model must:

• Hold the entire document in its context window
• Track hundreds of items simultaneously
• Maintain attention across thousands of tokens
• Avoid confusing similar items that appear pages apart

This leads to common failure modes: missed items, values assigned to wrong rows, and degraded accuracy toward the end of long documents.

Parallel Extraction solves this by providing focused, relevant context for each item:

Aspect	Standard Extraction	Parallel Extraction
Context per item	Entire document (all pages)	Only the relevant segment
Noise level	High (hundreds of unrelated items)	Minimal (just the target item)
Attention dilution	Significant on long documents	None — laser-focused extraction
Position bias	Later items often less accurate	Equal accuracy for all items

By cropping the document to show only the region containing each specific item, the LLM can dedicate its full attention and reasoning capacity to extracting that single item correctly. This is the same principle behind RAG (Retrieval-Augmented Generation) — less noise, more signal, better results.

When to Use Parallel Extraction

Parallel Extraction is ideal when:

• Your schema contains arrays of objects (e.g., line_items: [{sku, description, quantity, price}])
• Documents have many repeating items (10+ items benefit most)
• You need faster turnaround on large documents
• Items can be uniquely identified by a key field

Usage

Enable Parallel Extraction by specifying the parallel_ocr_keys parameter in your extraction request:

from retab import Retab

client = Retab()

response = client.documents.extract(
    document="invoice.pdf",
    json_schema=my_schema,
    parallel_ocr_keys={
        "line_items": "product_name"  # parent_path: child_key_path
    }
)

The parallel_ocr_keys parameter is a dictionary mapping:

• Key: The path to the array in your schema (e.g., "line_items", "transactions", "items.products")
• Value: The field within each array item that uniquely identifies it (e.g., "product_name", "sku", "id")

Example Schema

{
  "type": "object",
  "properties": {
    "invoice_number": { "type": "string" },
    "date": { "type": "string" },
    "line_items": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "product_name": { "type": "string" },
          "quantity": { "type": "number" },
          "unit_price": { "type": "number" },
          "total": { "type": "number" }
        }
      }
    },
    "total_amount": { "type": "number" }
  }
}

With parallel_ocr_keys={"line_items": "product_name"}, Retab will:

1. Extract all product names from the document
2. Locate each product’s position in the document
3. Extract each line item’s details in parallel
4. Merge results and extract constants (invoice_number, date, total_amount) separately

Supported Document Types

• PDF documents
• Images (JPEG, PNG, etc.)
• Office documents (DOCX, PPTX, ODT, ODP)
• Excel spreadsheets (XLSX, XLS)

Performance Benefits

Document Size	Standard Extraction	Parallel Extraction
5 line items	~3s	~3s
20 line items	~8s	~4s
50 line items	~15s	~5s
100+ items	~30s+	~6s

Times are approximate and vary based on document complexity and model used.

Consensus with Parallel Extraction

Parallel Extraction fully supports the n_consensus parameter. When enabled, each item is extracted multiple times independently, and results are compared to improve accuracy. This is particularly useful for:

• High-value documents requiring verification
• Documents with challenging handwriting or scan quality
• Compliance-critical extractions

response = client.documents.extract(
    document="invoice.pdf",
    json_schema=my_schema,
    parallel_ocr_keys={"line_items": "sku"},
    n_consensus=3  # Each item extracted 3 times for verification
)

Pricing

Parallel Extraction uses the same credit-based pricing as standard extraction. The cost is calculated per page:

credits/page = n_consensus × model_credits

Additionally, Parallel Extraction includes a key discovery pass that scans the document to identify and locate all items — this adds one extra extraction call at the document level.

Cost Breakdown

Component	Cost
Key discovery	1 × model_credits × page_count
Per-item extraction	n_items × n_consensus × model_credits
Constants extraction	1 × model_credits × page_count

Example

For a 10-page invoice with 25 line items using auto-small (1.0 credit) and n_consensus=1:

• Key discovery: 1 × 1.0 × 10 = 10 credits
• Item extraction: 25 × 1 × 1.0 = 25 credits
• Constants: 1 × 1.0 × 10 = 10 credits
• Total: 45 credits

While Parallel Extraction may cost more than standard extraction for the same document, the improved accuracy and speed often provide better value — especially when re-extractions due to errors are factored in.

For detailed pricing information, see the Pricing documentation.