Skip to main content
Complete sift-kg pipeline output from the unsealed Giuffre v. Maxwell court deposition, demonstrating OCR extraction from a scanned PDF and producing a graph of 190 entities and 387 relations.

​
Overview

This example demonstrates knowledge graph extraction from legal documents with OCR-scanned content, showing how sift-kg handles complex PDFs where text must be extracted via optical character recognition.

View Interactive Graph

Open examples/epstein/output/graph.html in your browser

Source Document

1 PDF with 36 sections (unsealed court deposition)

​
Quick Start

# View the interactive graph (no installation needed)
open examples/epstein/output/graph.html     # macOS
xdg-open examples/epstein/output/graph.html # Linux

# Or use sift's built-in viewer
sift view -o examples/epstein/output

​
Pipeline Output

The output/ directory contains the complete pipeline results: 
epstein/
├── docs/                          # Source PDF (unsealed court deposition)
└── output/
    ├── extractions/               # Per-document entity+relation JSON from LLM
    ├── graph_data.json            # Knowledge graph (226 entities, 708 relations)
    ├── merge_proposals.yaml       # Entity merge decisions (16 confirmed merges)
    ├── relation_review.yaml       # Flagged relations reviewed
    ├── communities.json           # Detected graph communities
    ├── entity_descriptions.json   # AI-generated entity descriptions
    ├── narrative.md               # Prose narrative with entity profiles
    └── graph.html                 # Interactive pyvis graph viewer

​
Pipeline Statistics

MetricValue
Input Document1 PDF, 36 sections
Document TypeUnsealed court deposition (OCR-scanned)
Initial Entities226 entities, 708 relations (after build + postprocess)
Entity Types93 persons, 56 locations, 24 organizations, 15 events, 2 vehicles
After Resolution190 entities, 387 relations (16 entity merges confirmed)
Model Usedclaude-haiku-4-5-20251001

​
OCR Processing

This example requires the --ocr flag during extraction to handle scanned PDFs: 
sift extract examples/epstein/docs --model openai/gpt-4o-mini -o my-output --ocr
The --ocr flag:
  • Uses optical character recognition to extract text from scanned images
  • Handles PDFs where text is not digitally encoded
  • Works with low-quality scans, handwritten documents, and mixed content
  • Automatically falls back to OCR when native text extraction fails
OCR processing may increase extraction time and cost. Ensure your documents actually require OCR before using this flag.

​
Entity Types

The deposition primarily contains:
  • Persons (93) — Named individuals mentioned in testimony
  • Locations (56) — Places referenced in the deposition
  • Organizations (24) — Companies, institutions, and groups
  • Events (15) — Specific incidents and occasions
  • Vehicles (2) — Named aircraft or vehicles
These entity types are automatically detected based on the content — no custom domain configuration was needed.

​
Resolution and Review

This example includes:

​
Entity Merges

16 entity merges were confirmed via sift resolve and human review: 
merges:
  - primary: "Jeffrey Epstein"
    duplicates:
      - "Epstein"
      - "Mr. Epstein"
    status: CONFIRMED

​
Relation Review

The relation_review.yaml file contains flagged relations that were reviewed for accuracy: 
relations:
  - source: "Person A"
    target: "Location B"
    relation_type: "VISITED"
    status: CONFIRMED
    notes: "Explicitly stated in testimony"
This demonstrates sift-kg’s support for reviewing potentially sensitive or ambiguous relationships.

​
Key Insights from the Graph

The generated narrative (narrative.md) provides:
  • Overview — High-level summary of testimony content
  • Entity descriptions — AI-generated profiles for key persons, locations, and organizations
  • Community detection — Clustered subgraphs showing relationship networks
This example demonstrates technical capabilities for processing legal documents. The content itself is a matter of public record from unsealed court proceedings.

​
Re-running the Example

​
Option 1: Build from Existing Extractions (Free)

Use the pre-extracted entities — no LLM API calls: 
pip install sift-kg
sift build -o examples/epstein/output
sift view -o examples/epstein/output

​
Option 2: Full Pipeline from Scratch

Re-extract entities from the source PDF with OCR: 
# Extract entities with OCR support
sift extract examples/epstein/docs \
  --model openai/gpt-4o-mini \
  -o my-output \
  --ocr

# Build the knowledge graph
sift build -o my-output

# Resolve duplicate entities
sift resolve -o my-output --model openai/gpt-4o-mini

# Review proposed merges
sift review -o my-output

# Apply confirmed merges
sift apply-merges -o my-output

# Generate narrative
sift narrate -o my-output --model openai/gpt-4o-mini

# View the result
sift view -o my-output
Always use the --ocr flag when working with scanned documents, court filings, or any PDF where you cannot select and copy text.

​
Use Cases

This example pattern works well for:
  • Legal discovery — Map entities and relationships from depositions and court documents
  • Investigative research — Extract structured data from public records
  • Historical archives — Process scanned documents and old records
  • Due diligence — Analyze legal filings and testimony

​
Processing Scanned Documents

When working with OCR documents:
  1. Test OCR quality — Preview extracted text to ensure readability
  2. Adjust models — Consider using more powerful models (GPT-4) for complex layouts
  3. Review carefully — OCR errors can propagate to entity extraction
  4. Use resolution — Merge entities that may have OCR-induced variations

​
Next Steps

Explore Other Examples

See Transformers and FTX examples

OCR Guide

Learn more about processing scanned documents

Build docs developers (and LLMs) love

Get started for freeTalk to us