Document Processing

Document processing is the discipline of converting raw, unstructured documents (PDFs, scanned images, invoices, medical records, legal contracts) into that machines can work with. It encompasses the full pipeline from ingestion through extraction to downstream consumption.

A critical distinction: is a component of document processing, not a synonym for it. A complete document processing system handles:

• Spatial layout analysis: identifying columns, headers, footers, sidebars, and reading order
• Semantic structure extraction: distinguishing titles from body text, captions from data
• Table detection and parsing: recovering row-column relationships
• Entity recognition: identifying dates, monetary amounts, names, and domain-specific fields
• Character recognition (OCR): converting pixel patterns into text characters

ETL Pipelines for Documents

Production document systems follow the pattern, a standard data engineering framework applied to unstructured content:

OCR at Scale: The Structural Bottleneck

Traditional (Optical Character Recognition) converts pixel patterns into character sequences. While effective for simple text recognition, it introduces two critical bottlenecks in production systems:

1. Structural Information Loss

When OCR flattens a document into a character stream, spatial relationships are discarded. Tables lose row-column associations, headers lose hierarchical context, and multi-column layouts become interleaved text. Downstream systems must then reconstruct structure from ambiguous signals.

2. Token Explosion

Text converts OCR output into discrete symbols for language model consumption. Complex documents routinely produce thousands of per page, much of which encodes redundant structural information rather than semantic content.

LLM-Based OCR

Large models (GPT-4V, Claude, Gemini) can process document images directly, extracting fields, answering questions, and reasoning about layout. However, they serve a fundamentally different role than dedicated OCR systems:

Contexts Optical Compression

DeepSeek-OCR introduces a paradigm called Contexts Optical Compression, a fundamental inversion of the traditional OCR approach:

• Traditional OCR: Image → Character sequence → Text tokens (many tokens, structure lost)
• DeepSeek-OCR: Image → Visual → Compressed vision tokens (few tokens, structure preserved)

The core insight: a page of text rendered as an image can be represented with far fewer vision tokens than the equivalent text tokens. By compressing at the visual representation level rather than the text level, both token count and spatial structure are preserved simultaneously.

This reframes OCR from a recognition problem (pixels to characters) into a problem (visual information to minimal tokens). It doesn't replace LLM reasoning. It makes downstream systems cheaper, faster, and scalable.

Architecture

DeepSeek-OCR consists of two components: a that compresses document images into compact token sequences, and a language decoder that generates structured text output.

DeepEncoder: Dual-Pathway Vision

The encoder combines two specialized vision models in parallel:

• -base (80M params): window attention for fine-grained visual perception at patch-size 16
• -large (300M params): dense global attention for high-level semantic understanding

A 2-layer convolutional module between them performs 16x . This is where the primary token compression occurs.

MoE Decoder

The decoder uses a architecture: 3B total parameters across 64 experts, but only 6 experts + 2 shared are activated per token (~570M active parameters). This delivers 3B-class output quality at sub-600M inference cost.

Resolution Modes

DeepSeek-OCR 2: Visual Causal Flow

Released January 2026, OCR v2 replaces the encoder architecture entirely:

• Qwen2-0.5B replaces CLIP: shifts from spatial scanning to semantic language-model-based reasoning
• Visual Causal Flow: cascaded 1D causal structures achieve 2D spatial reasoning. Visual tokens use bidirectional attention; query tokens use causal triangular attention
• Dynamic semantic reordering: reads content by logical structure (title → body → tables) rather than rigid left-to-right, top-to-bottom scanning

Cost & Throughput

Vision-based compression directly impacts three cost dimensions:

• cost: 15–26x fewer tokens per page reduces downstream LLM API spend proportionally
• : single A100-40G (40 GB ) processes 200,000+ pages/day; a 20-node cluster handles ~33M pages/day
• Quality-adjusted cost: preserved layout reduces post-processing errors, lowering human review costs

Pipeline Integration

DeepSeek-OCR integrates into existing pipelines as an Transform-stage replacement or augmentation. Standard upstream (ingestion) and downstream (entity extraction, validation, loading) steps remain unchanged.

Hybrid Routing Pattern

Production systems route documents to different OCR backends based on complexity, cost tolerance, and accuracy requirements:

Tool Comparison

DeepSeek-OCR + + LLM is an emerging production pattern: DeepSeek for ingestion/compression, ColPali for retrieval, LLMs for reasoning. Each model applied where it provides maximum leverage.

Limitations