HeritageNet Example

This repo builds HeritageNet, a structured knowledge graph from historical medical texts, using large language models and Neo4j.

Files

• main.py — Loads text, extracts graph data, and writes it to Neo4j
• KGAgents.py — Defines the KnowledgeGraphAgent using the camel-ai library
• sample_input.txt — Contains 20 randomly sampled pages from:
  Traité de la moelle épinière et de ses maladies, Ollivier (1827)
• output_viz.html — Optional visualization of the resulting graph

Installation

pip install "camel-ai[all]==0.2.16" neo4j
export GROQ_API_KEY="your_key"

Usage

Configure Neo4j credentials in main.py, then run:

python main.py

HeritageNet Taxonomy

HeritageNet captures medical knowledge as reported in traditional and early modern texts, without modern reinterpretation.

Core Evidence Entities

ClinicalObservation
Historical signs, symptoms, or disease states e.g. “congestions sanguines rachidiennes”

TherapeuticOutcome
Documented effects of treatments (including failures or side effects)
e.g. convulsions

ContextualFactor
Environmental, demographic, or behavioral correlates
e.g. suppressed sweat, lochia

MechanisticConcept
Explanatory theories proposed at the time
e.g. spinal vein congestion causes motor paralysis

TherapeuticApproach
Treatments and their preparation/administration
e.g. hydrocyanic acid on vesicatory wounds

SourceText
Bibliographic origin of the knowledge
e.g. Ollivier’s 1824 treatise

Evidence Relationships

• co_occurs_with: Joint appearance of entities
• preceded_by, followed_by: Temporal flow
• modified_by: Changes due to external/internal factors
• responds_to: Symptom change after treatment
• associated_with: Loose historical association
• results_in: Direct effects
• described_in: Tied to source text
• contradicts, corroborates: Consistency/conflict between items

Each relation can include metadata qualifiers to preserve historical ambiguity.

---

Project

This toy example is part of Chronos. Chronos mines overlooked historical texts to generate novel, testable hypotheses using LLMs and structured graph reasoning.