Document Graph Extractor

The document graph extractor processes PDFs by extracting spatial blocks with layout classification (paragraphs, tables, forms, lists, headers, footers, figures, handwriting). Includes confidence scoring and optional VLM correction for low-confidence blocks. Best for archival documents, scanned files, and documents requiring spatial understanding.

View extractor details at api.mixpeek.com/v1/collections/features/extractors/document_graph_extractor_v1 or fetch programmatically with GET /v1/collections/features/extractors/{feature_extractor_id}.

Pipeline Steps

Filter Dataset (if collection_id provided)
- Filter to specified collection
PDF URL Resolution
- Find PDF URL from row data (data, pdf_url, document_url, file_url, etc.)
- Convert S3 keys to full S3 URLs if needed
Layout Detection Mode Fork
- If use_layout_detection=true (NEW - ML-based): a. PaddleOCR layout detection (finds ALL elements: text, images, tables) b. Skip to Step 4 (object_type already set by detector)
- If use_layout_detection=false (LEGACY - Text-only): a. PyMuPDF span extraction (text with bounding boxes) b. Spatial clustering (group nearby spans into logical blocks) c. Layout classification (rule-based: paragraph, table, form, etc.)
Confidence Scoring
- Score extraction quality with A/B/C/D tags
- Based on OCR quality, spatial coherence, text patterns
Text Cleaning
- Remove OCR artifacts
- Normalize whitespace
Page Rendering (conditional: if generate_thumbnails=true OR use_vlm_correction=true)
- Full page thumbnails at configured DPI
- Segment-level thumbnails for each block
VLM Correction (conditional: if use_vlm_correction=true AND NOT fast_mode AND confidence C/D)
- Gemini/OpenAI/Anthropic vision models correct low-confidence text
- Only applied to blocks with poor extraction quality
Text Embedding (conditional: if run_text_embedding=true)
- E5-Large embeddings (1024D) for semantic search
Output
- Block-level documents with text, layout type, bbox, confidence, embeddings

When to Use

Use Case	Description
Archival documents	Extract structured data from scanned historical documents
Scanned PDFs	Process documents with mixed text quality
Forms processing	Identify and extract form fields, tables, and structured data
Document understanding	Analyze document layout and structure
Spatial search	Find specific sections or blocks within documents
Multi-layout documents	Process documents with complex layouts (reports, contracts, etc.)

When NOT to Use

Scenario	Recommended Alternative
Simple text extraction	`text_extractor`
Images only	`image_extractor`
Video/audio content	`multimodal_extractor`
Born-digital PDFs with perfect text	`text_extractor` (faster, simpler)

Input Schema

Field	Type	Required	Description
`pdf`	string	Yes	URL or S3 path to PDF file. Supports multi-page PDFs.

{
  "pdf": "s3://my-bucket/documents/invoice-2024.pdf"
}

Input Examples:

Type	Example
Invoice	`s3://documents/invoices/inv-001.pdf`
Contract	`https://cdn.example.com/contracts/lease.pdf`
Scanned document	`s3://archive/scanned/1985-report.pdf`
Form	`s3://forms/application-form.pdf`

Supported Formats: PDF only Recommended: 150-300 DPI for scanned documents Max File Size: 100MB per PDF

Output Schema

Each spatial block produces one document with the following fields:

Field	Type	Description
`text`	string	Extracted text content (raw or VLM-corrected)
`object_type`	string	Layout type: paragraph, table, form, list, header, footer, figure, handwritten
`bbox`	object	Bounding box `{x, y, width, height}` in PDF coordinates
`page_number`	integer	Page number (0-indexed)
`confidence_tag`	string	Confidence grade: A (high), B (good), C (fair), D (poor)
`confidence_score`	number	Confidence score (0.0-1.0)
`document_graph_extractor_v1_text_embedding`	float[1024]	E5-Large text embedding (if enabled)
`page_image_url`	string	Full page thumbnail URL (if generated)
`segment_thumbnail_url`	string	Block-specific thumbnail URL (if generated)
`thumbnail_url`	string	Page thumbnail URL (if generated)

{
  "text": "INVOICE #12345\nDate: January 15, 2024\nAmount Due: $1,250.00",
  "object_type": "header",
  "bbox": {"x": 50, "y": 720, "width": 500, "height": 80},
  "page_number": 0,
  "confidence_tag": "A",
  "confidence_score": 0.95,
  "document_graph_extractor_v1_text_embedding": [0.023, -0.041, ...],
  "page_image_url": "s3://mixpeek/thumbnails/page_0.jpg",
  "segment_thumbnail_url": "s3://mixpeek/thumbnails/seg_0_header.jpg"
}

Parameters

Layout Detection Parameters

Parameter	Type	Default	Description
`use_layout_detection`	boolean	`false`	Use ML-based PaddleOCR layout detection (finds images + tables + text) vs legacy text-only extraction
`render_dpi`	integer	`150`	DPI for PDF page rendering (72-300). Higher = better quality, slower processing

VLM Correction Parameters

Parameter	Type	Default	Description
`use_vlm_correction`	boolean	`false`	Enable VLM correction for low-confidence blocks (C/D tags)
`min_confidence_for_vlm`	string	`"C"`	Minimum confidence tag to trigger VLM correction: A, B, C, or D
`vlm_provider`	string	`"google"`	VLM provider: `google`, `openai`, `anthropic`
`vlm_model`	string	`"gemini-3.1-flash-lite"`	Specific VLM model for correction
`fast_mode`	boolean	`false`	Skip VLM correction even if enabled (for faster processing)

Clustering Parameters (Legacy mode only)

Parameter	Type	Default	Description
`vertical_threshold`	number	`10.0`	Vertical distance threshold for grouping text spans
`horizontal_threshold`	number	`5.0`	Horizontal distance threshold for grouping text spans
`min_text_length`	integer	`1`	Minimum text length to include in blocks

Confidence & Embedding Parameters

Parameter	Type	Default	Description
`base_confidence`	number	`0.8`	Base confidence score for extracted blocks
`run_text_embedding`	boolean	`true`	Generate E5-Large embeddings for semantic search

Thumbnail Parameters

Parameter	Type	Default	Description
`generate_thumbnails`	boolean	`true`	Generate page and segment thumbnails
`thumbnail_dpi`	integer	`72`	DPI for thumbnail generation
`thumbnail_mode`	string	`"fit"`	Thumbnail resize mode: `fit`, `fill`, `crop`

Configuration Examples

{
  "feature_extractor": {
    "feature_extractor_name": "document_graph_extractor",
    "version": "v1",
    "input_mappings": {
      "pdf": "payload.document_url"
    },
    "field_passthrough": [
      { "source_path": "metadata.invoice_id" },
      { "source_path": "metadata.vendor" }
    ],
    "parameters": {
      "use_layout_detection": true,
      "render_dpi": 150,
      "generate_thumbnails": true,
      "run_text_embedding": true
    }
  }
}

Performance & Costs

Metric	Value
Processing speed	~1-5 pages/sec (depends on DPI and features enabled)
Layout detection	~500ms per page (PaddleOCR)
VLM correction	~2s per low-confidence block
Embedding generation	~5ms per block
Cost (minimal)	~$0.001/page (text extraction only)
Cost (with VLM)	~ $0.01-$ 0.05/page (depends on # of low-confidence blocks)

Vector Index

Property	Value
Index name	`document_graph_extractor_v1_text_embedding`
Dimensions	1024
Type	Dense
Distance metric	Cosine
Datatype	float32
Inference model	`multilingual_e5_large_instruct_v1`

Layout Types

The extractor classifies blocks into these layout types:

Type	Description	Example Use Case
`paragraph`	Body text blocks	Article content, descriptions
`table`	Tabular data	Financial tables, data grids
`form`	Form fields and labels	Application forms, surveys
`list`	Bulleted or numbered lists	Requirements, instructions
`header`	Page headers	Document titles, section headers
`footer`	Page footers	Page numbers, disclaimers
`figure`	Images and captions	Charts, diagrams, photos
`handwritten`	Handwritten text	Signatures, annotations

Confidence Tags

Extraction quality is graded with confidence tags:

Tag	Confidence	Description	Action
A	0.9-1.0	Excellent	No correction needed
B	0.7-0.9	Good	Minor issues, usually acceptable
C	0.5-0.7	Fair	VLM correction recommended
D	0.0-0.5	Poor	VLM correction strongly recommended

Comparison: ML Layout Detection vs Legacy

Feature	ML Layout Detection	Legacy Text-Only
Finds images	✅ Yes	❌ No
Finds tables	✅ Yes (better accuracy)	⚠️ Basic heuristics
Processing speed	Slower (~500ms/page)	Faster (~100ms/page)
Best for	Complex layouts, scanned docs	Simple text-only PDFs
Model	PaddleOCR	PyMuPDF + heuristics

Limitations

PDF only: Does not process images, Word docs, or other formats
Memory intensive: Large PDFs (100+ pages) may require increased memory
VLM costs: VLM correction adds significant cost for low-confidence documents
Language support: OCR works best with Latin scripts; non-Latin may have reduced accuracy
Handwriting: Handwritten text detection is experimental and less reliable

Get Started

What Mixpeek Extracts

Retrieval

Platform

Resources

Document Graph Extractor

Pipeline Steps

When to Use

When NOT to Use

Input Schema

Output Schema

Parameters

Layout Detection Parameters

VLM Correction Parameters

Clustering Parameters (Legacy mode only)

Confidence & Embedding Parameters

Thumbnail Parameters

Configuration Examples

Performance & Costs

Vector Index

Layout Types

Confidence Tags

Comparison: ML Layout Detection vs Legacy

Limitations

Get Started

What Mixpeek Extracts

Retrieval

Platform

Resources

Documentation Index

​Pipeline Steps

​When to Use

​When NOT to Use

​Input Schema

​Output Schema

​Parameters

​Layout Detection Parameters

​VLM Correction Parameters

​Clustering Parameters (Legacy mode only)

​Confidence & Embedding Parameters

​Thumbnail Parameters

​Configuration Examples

​Performance & Costs

​Vector Index

​Layout Types

​Confidence Tags

​Comparison: ML Layout Detection vs Legacy

​Limitations

​Related

Pipeline Steps

When to Use

When NOT to Use

Input Schema

Output Schema

Parameters

Layout Detection Parameters

VLM Correction Parameters

Clustering Parameters (Legacy mode only)

Confidence & Embedding Parameters

Thumbnail Parameters

Configuration Examples

Performance & Costs

Vector Index

Layout Types

Confidence Tags

Comparison: ML Layout Detection vs Legacy

Limitations

Related