SNCF Safety Knowledge Extraction Pipeline

AI TINKERERS Hackaton 10/12/25

This project extracts structured railway safety knowledge from unstructured incident documents and prepares it for ingestion into a Neo4J graph-based safety system.
Its purpose is to help railway safety personnel analyze incidents, identify root causes, and justify maintenance or budget decisions using a connected safety knowledge graph. The data is from the SNCF opendata and can be viewed here: https://eu.ftp.opendatasoft.com/sncf/rapports/rapport-securite-2023.PDF This a basic version of the graph, it should be enriched to demonstrate its full capability.

---

Overview

1. PDF → Markdown

Documents are converted to markdown using Docling. This provides a uniform text format for analysis.

2. Markdown Chunking

The markdown document is segmented into sections based on titles (#, ##, etc.).
Each chunk is processed independently.

3. Three Classification Modules

A. Incident Classification

Outputs:

• incident_type
• possible_impacts
• possible_root_causes (with HIGH/LOW likelihood)
• Cypher defining:
  • Incident → Impact
  • Incident → RootCause
  • Impact → RootCause

B. Cause Classification

Outputs:

• root cause (chosen from a fixed taxonomy)
• normalized causal category
• Cypher linking:
  • Category → RootCause

C. Impact Classification

Outputs:

• impact category
• normalized sub-impact subtype
• optional inferred root cause
• Cypher linking:
  • SubImpact → Impact

4. JSON Outputs

Each classifier produces a JSON file:

├── incidents.json ├── causes.json └── impacts.json

All files are ready for Neo4J ingestion.

---

FastAPI Endpoints

Convert PDF to Markdown

POST /convert_pdf?pdf_path=path/to/file.pdf

Process Incidents

POST /process_incidents?md_path=path/to/file.md

Process Causes

POST /process_causes?md_path=path/to/file.md

Process Impacts

POST /process_impacts?md_path=path/to/file.md

Neo4J Integration

The pipeline produces structured JSON files (incidents.json, causes.json, impacts.json) that already contain fully generated Cypher statements.
These Cypher statements express all required graph relationships for a safety-knowledge model.

What Gets Written to Neo4J

The system generates and ingests four major types of nodes:

• IncidentType
• Impact
• SubImpact
• RootCause
• SubCause

And several relationship types:

• (:IncidentType)-[:HAS_IMPACT]->(:Impact)
• (:IncidentType)-[:HAS_CAUSE {likelihood}]->(:RootCause)
• (:Impact)-[:LINKED_TO_CAUSE]->(:RootCause)
• (:SubImpact)-[:EST_SOUS_IMPACT_DE]->(:Impact)

These relations form a multi-layered safety graph connecting:

Incident → Impact → Cause

This creates a navigable structure where analysts can:

• explore causal chains
• compare incident families
• trace common failure patterns
• justify maintenance decisions
• support risk scoring and prioritization

How Ingestion Works

The ingestion route (/neo4j_ingest) reads the three JSON files and extracts every Cypher statement.

driver.execute_query(cypher)