dimeai 0.1.0

DIMEFIL Analyst CLI - AI-powered geopolitical event analysis using temporal graph networks

DimeAI

DIMEFIL Analyst CLI - AI-powered geopolitical event analysis.

Installation

cd dimeai
uv pip install -e .

Quick Start

# Collect articles about a situation
dimeai collect "South China Sea tensions 2024" -n 20

# Extract entities and relationships
dimeai extract -i articles -o graph.json

# Train TGN model
dimeai train -g graph.json -o model.pt

# Analyze a situation
dimeai analyze "China coast guard incident" -m model.pt -g graph.json

# Interactive session
dimeai interactive -g graph.json

# Run benchmarks
dimeai benchmark -g graph.json

Commands

Command	Description
collect	Collect articles via web scraping
extract	Extract entities with GLiNER 2
train	Train TGN model
analyze	Analyze a situation
simulate	Interactive what-if simulation
interactive	REPL for graph exploration
benchmark	Compare TGN vs baselines
features compute	Compute semantic features
features inspect	Inspect node features
config show	Show configuration
config set	Set configuration value
dossier list	List investigation dossiers
dossier create	Create new dossier
dossier show	Show dossier details
dossier open	Interactive dossier session
dossier collect	Add articles to dossier
dossier export	Export dossier as graph
dossier note	Add analyst note
dossier delete	Delete dossier

Dossier System

Dossiers are persistent investigation sessions stored in SQLite (~/.dimeai/dossiers.db).

# Create a new dossier
dimeai dossier create "SCS Tensions 2024" -d "Investigating recent incidents"

# Collect articles (uses DuckDuckGo search)
dimeai dossier collect 1 "Philippines coast guard incident" -n 5

# Open interactive session
dimeai dossier open 1

# In session:
#   collect <query>  - Search and add articles
#   extract          - Extract entities from articles
#   articles         - List collected articles
#   entities         - List extracted entities
#   events           - List events
#   note <text>      - Add observation
#   hypothesis <text> - Add hypothesis
#   analyze          - Analyze patterns
#   export           - Export as graph JSON
#   quit             - Exit session

# Export dossier
dimeai dossier export 1 -o my_investigation.json

DIMEFIL Framework

• Diplomatic - Protests, talks, negotiations
• Information - Propaganda, disinformation
• Military - Exercises, deployments
• Economic - Fishing, trade, sanctions
• Financial - (not yet implemented)
• Intelligence - (not yet implemented)
• Law Enforcement - Coast guard, patrols

Model Performance

Model	Accuracy	Macro F1	vs Random F1
Random	16.0%	14.1%	-
Majority	20.1%	5.6%	-8.5pp
TGN	28.7%	20.9%	+6.9pp

TGN beats both baselines on Macro F1 (the right metric for imbalanced data). Class-weighted loss helps the model learn rare classes (diplomatic, information).

Limitations

1. Low absolute accuracy - 26.6% on 7-class problem
2. Class imbalance - legal (36%), law_enforcement (13%)
3. Temporal patterns - Response patterns show 79% law_enforcement → law_enforcement
4. Graph heuristics - Agent uses heuristics, not neural predictions

Development

# Run tests
python -m pytest dimeai/tests/ -v

# Run property-based tests
python -m pytest dimeai/tests/test_properties.py -v