Knowledge-Graph-Agent

Knowledge Graph Agent

A 5-node LangGraph agent that extracts entities and relationships from any domain document and builds a queryable knowledge graph with interactive visualization.

Architecture

Document Ingester → Entity Extractor → Relationship Extractor → Graph Builder → Graph Summarizer

Node	Responsibility
Document Ingester	Parse text / fetch URL, clean and chunk into 1500-char segments
Entity Extractor	LLM JSON extraction per chunk, deduplication, ID assignment
Relationship Extractor	LLM JSON extraction against known entity list, deduplication
Graph Builder	Build NetworkX DiGraph, compute density/centrality/components, store in-memory
Graph Summarizer	Stream a 3–5 paragraph natural language summary of the graph

Separate Query Engine (core/query_engine.py)

Query Type	Trigger	Algorithm
Neighborhood	1 entity name in query	nx.ego_graph(radius=2)
Path Finding	2+ entity names in query	nx.shortest_path on undirected view
Keyword Search	No entity names matched	Term scoring on entity names + descriptions

Stack

• Backend: FastAPI + LangGraph + NetworkX + Ollama (llama3.2)
• Frontend: Streamlit + pyvis (interactive graph)
• Storage: In-memory dict[str, nx.DiGraph] (graph store)

Entity & Relationship Types

Entity types: PERSON · ORGANIZATION · TECHNOLOGY · CONCEPT · LOCATION · EVENT · PRODUCT · PROCESS

Relationship types: WORKS_FOR · PART_OF · USES · CAUSES · IS_A · RELATED_TO · COMPETES_WITH · LOCATED_IN · DEVELOPS · FOUNDED_BY · LEADS · OWNS · COLLABORATES_WITH · DEPENDS_ON · PRECEDES · FOLLOWS

Quick Start

ollama pull llama3.2

# Backend
cd backend && pip install -r requirements.txt
uvicorn main:app --reload

# Frontend (new terminal)
cd frontend && pip install -r requirements.txt
streamlit run app.py

Or via Make:

make install-backend install-frontend
make run-backend    # terminal 1
make run-frontend   # terminal 2

Docker

docker compose up --build -d
docker exec kga-ollama ollama pull llama3.2

Endpoints

Method	Path	Description
GET	/health	System health check
POST	/api/graph/build	Blocking graph build
POST	/api/graph/build/stream	SSE streaming build
GET	/api/graph	List all built graphs
GET	/api/graph/{id}	Get graph data
GET	/api/graph/{id}/export	Export full graph as JSON
POST	/api/graph/query	Semantic graph query

Frontend Tabs

• Build Graph — Document input + 5-agent streaming progress + live summary
• Explore Graph — Entity table (filterable by type) + relationship table + interactive pyvis network + downloads
• Query Graph — Natural language query → auto-detected query type → sub-graph visualization + LLM answer

Config

Variable	Default	Description
CHUNK_SIZE	1500	Characters per text chunk
MAX_CHUNKS	6	Max chunks processed per document
MAX_ENTITIES_PER_CHUNK	10	Max entities extracted per chunk
MAX_RELS_PER_CHUNK	8	Max relationships extracted per chunk
MAX_GRAPH_NODES	200	Global entity cap