FTX Collapse Example

Knowledge graph from 9 Wikipedia articles about the FTX cryptocurrency exchange collapse

Complete sift-kg pipeline output from 9 Wikipedia articles documenting the FTX cryptocurrency exchange collapse, producing a graph of 373 entities and 1,184 relations after deduplication.

Overview

This example demonstrates knowledge graph extraction from journalistic and encyclopedic content, with emphasis on the entity resolution workflow to merge duplicate entities across multiple documents.

View Interactive Graph

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

Source Documents

9 text files covering FTX, Alameda Research, Binance, and key people

Quick Start

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

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

Pipeline Output

The output/ directory contains the complete pipeline results:

ftx/
├── docs/                          # 9 source documents (~148K total)
└── output/
    ├── extractions/               # Per-document entity+relation JSON from LLM
    ├── graph_data.json            # Knowledge graph (373 entities, 1184 relations)
    ├── merge_proposals.yaml       # Entity merge decisions (CONFIRMED/REJECTED)
    ├── entity_descriptions.json   # AI-generated entity descriptions
    ├── narrative.md               # Prose narrative with entity profiles
    └── graph.html                 # Interactive pyvis graph viewer

Pipeline Statistics

Metric	Value
Input Documents	9 text files (~148K total)
Topics Covered	FTX, Alameda Research, Binance, Sam Bankman-Fried, and other key figures
Raw Entities Extracted	~777 entities from LLM
After Pre-dedup (semhash)	750 entities (27 deterministic merges)
After Build + Postprocess	432 entities, 1,201 relations
After Resolution (3 passes)	373 entities, 1,184 relations (59 entities merged via LLM + human review)
Final Entity Descriptions	100 entity profiles in narrative
Model Used	claude-haiku-4-5-20251001
Total Cost	~$0.28 (extraction was separate)

Entity Resolution Workflow

This example showcases the full deduplication pipeline:

1. Automatic Semantic Deduplication

During extraction, semantic hashing automatically merges near-identical entities:

• Before: 777 raw entities
• After: 750 entities (27 deterministic merges)

2. Build + Postprocess

Graph construction with normalization and filtering:

• Result: 432 entities, 1,201 relations

3. LLM-Assisted Resolution

Three passes of sift resolve to identify remaining duplicates:

sift resolve -o examples/ftx/output --model claude-haiku-4-5-20251001

This generates merge_proposals.yaml with candidate merges:

merges:
  - primary: "Sam Bankman-Fried"
    duplicates:
      - "SBF"
      - "Samuel Bankman-Fried"
    status: CONFIRMED
    reasoning: "Same person, different name variations"
  
  - primary: "FTX Trading Ltd."
    duplicates:
      - "FTX"
      - "FTX.com"
    status: CONFIRMED
    reasoning: "Same exchange entity"

4. Human Review

sift review -o examples/ftx/output

Interactive review to confirm or reject each proposed merge.

5. Apply Merges

sift apply-merges -o examples/ftx/output

Final result: 373 entities, 1,184 relations (59 entities merged)

Key Insights from the Graph

The generated narrative (narrative.md) provides:

• Overview — High-level synthesis of the FTX collapse timeline
• Entity descriptions — AI-generated profiles for 100 key entities (companies, people, events)
• Relationship mapping — How FTX, Alameda, Binance, and key figures are connected

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/ftx/output
sift view -o examples/ftx/output

Option 2: Full Pipeline from Scratch

Re-extract entities from the source documents:

# Extract entities
sift extract examples/ftx/docs \
  --model openai/gpt-4o-mini \
  -o my-output

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

# Resolve duplicate entities (3 passes recommended)
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

Run sift resolve multiple times (2-3 passes) to catch progressively more subtle duplicates. Each pass refines the merge proposals.

Cost Breakdown

The ~$0.28 total cost includes:

• Resolution — LLM calls across 3 passes to identify duplicates
• Narration — LLM calls to generate entity descriptions and overview

(Extraction cost was calculated separately.)

Costs will vary based on:

• Number of entities to resolve
• Model chosen (Haiku vs GPT-4o-mini vs others)
• Number of resolution passes

Use Cases

This example pattern works well for:

• Investigative journalism — Map connections across news articles
• Business intelligence — Track companies, people, and events across sources
• Historical analysis — Document timelines and relationships in major events
• Due diligence — Aggregate information about entities from multiple sources

Merge Proposals File

The merge_proposals.yaml file is a key artifact:

merges:
  - primary: "Primary Entity Name"
    duplicates:
      - "Duplicate 1"
      - "Duplicate 2"
    status: CONFIRMED  # or REJECTED
    reasoning: "Why these should be merged"

This file:

• Is generated by sift resolve
• Can be manually edited before applying
• Supports version control and collaboration
• Is applied with sift apply-merges

Next Steps

Explore Other Examples

See Transformers and Epstein examples

Resolution Guide

Learn more about entity deduplication