Mining Knowledge from Support Tickets — READY 2026 Demo

From READY 2026 Tech Exchange #42: AI-Powered Support: Mining Knowledge from Ticket Data

This repository demonstrates how semantic clustering turns a support ticket backlog into KB articles, and how IRIS GraphRAG enables hybrid vector + graph retrieval at scale.

Uses PlanetCare, a fictional EMR system — all data is synthetic, safe to share and adapt.

---

What's Here

Notebook	What it shows
planetcare_clustering_demo.ipynb	MDS gap → HDBSCAN clustering → resolution anchoring → KB article → wiki augmentation + Graph_KG audit
planetcare_system_demo.ipynb	iris-vector-graph API: kg_KNN_VEC, kg_VECTOR_GRAPH_SEARCH, kg_GRAPH_WALK, Cypher

Both notebooks require IRIS Community and an OpenAI API key (or compatible alternative).

---

Quick Start

Prerequisites:

• Docker (for IRIS Community — free container)
• OpenAI API key (or see .env.example for OpenRouter / Ollama alternatives)
• Python 3.10+

Step 1 — Clone and install

git clone https://github.com/intersystems-community/ready2026-knowledge-graph-demo
cd ready2026-knowledge-graph-demo
pip install -r requirements.txt

Step 2 — Start IRIS Community

cd docker && docker compose up -d && cd ..

First run pulls the IRIS image (~2 GB). Subsequent starts are fast.
IRIS is ready when docker ps shows healthy for planetcare-iris.

Step 3 — Set your API key

export OPENAI_API_KEY=sk-...
# Or: cp .env.example .env  (edit for OpenRouter, Ollama, etc.)

Step 4 — Initialize Graph_KG

python setup/setup_iris.py

This loads 276 PlanetCare tickets into IRIS Graph_KG, builds 384-dim embeddings via iris-vector-graph, and uses GPT-4o-mini to extract entity edges (AFFECTS / EXHIBITS / FIXED_BY). Takes ~5 minutes.

Expected output:

PlanetCare Demo — IRIS Setup
IRIS connection: OK
Embedder: all-MiniLM-L6-v2 (384-dim)
LLM: gpt-4o-mini
  276 ticket nodes created in Graph_KG
  276 embeddings stored
  606 graph edges created
Setup complete!

Step 5 — Open the notebooks

jupyter notebook notebooks/

Start with planetcare_clustering_demo.ipynb — that's the one that generated the most discussion at READY.

Using OpenRouter or Ollama instead? Copy .env.example to .env and set LLM_PROVIDER.
Any OpenAI-compatible endpoint works.

---

What You'll See

Clustering Demo (planetcare_clustering_demo.ipynb)

1. The data quality problem — 276 PlanetCare tickets, 58% with no resolution text. The norm, not the exception.
2. HDBSCAN clustering — embed problem descriptions with all-MiniLM-L6-v2, clusters emerge automatically without any labeling.
3. Why it works — the model encodes semantic intent, not keywords. Similar problems group together across different phrasings.
4. Resolution-anchored discovery — the 42% with solutions become anchors. Each anchor suggests fixes for similar unresolved tickets in its cluster.
5. KB article generation — GPT-4o-mini synthesizes a structured article from the cluster's anchor tickets.
6. Wiki augmentation — article appended to data/planetcare_wiki/billing.md (pre-existing doc with documented gaps). AUTHORED_KB and SOURCED_KB edges recorded in Graph_KG.

System Demo (planetcare_system_demo.ipynb)

Uses iris-vector-graph directly:

engine.kg_KNN_VEC(query_json, k=8)              # K-nearest by embedding similarity
engine.kg_VECTOR_GRAPH_SEARCH(query_json)        # hybrid: vector + graph expansion
ops.kg_GRAPH_WALK("pc_ticket:PC-00001", depth=2) # traverse entity edges
ops.kg_NEIGHBORHOOD_EXPANSION(entity_nodes)      # find tickets sharing same entities
engine.execute_cypher("MATCH (t:PCTicket)...")   # direct Cypher queries

---

Repo Contents

notebooks/
  planetcare_clustering_demo.ipynb   ← The main demo
  planetcare_system_demo.ipynb       ← iris-vector-graph API showcase

data/
  planetcare_demo_tickets.json       ← 276 synthetic PlanetCare tickets
  planetcare/
    questionnaire_clusters_anon.csv  ← 295 anonymized questionnaire tickets
    questionnaire_kb_articles_anon.json
  planetcare_wiki/
    billing.md                       ← Pre-existing KB, documented gaps
    laboratory.md                    ← Pre-existing KB, documented gaps
    pharmacy.md                      ← Mostly empty — gaps noted

setup/
  setup_iris.py    ← Initializes Graph_KG via IRISGraphEngine, loads tickets, builds embeddings
  embedder.py      ← Embedding provider abstraction (local / OpenAI / OpenRouter / custom)

docker/
  docker-compose.yml   ← IRIS Community container

.env.example         ← All configuration options
requirements.txt

---

Data

All data is synthetic — no real patient data, no real hospital names.

File	Description
data/planetcare_demo_tickets.json	276 synthetic PlanetCare tickets · 7 categories · 42% resolved / 58% open
data/planetcare/questionnaire_clusters_anon.csv	295 anonymized questionnaire tickets with cluster labels
data/planetcare_wiki/*.md	Pre-existing KB articles with documented knowledge gaps

---

PlanetCare Wiki

data/planetcare_wiki/ is the fictional PlanetCare KB.
Pre-existing articles document what's known, with explicit "Knowledge Gaps" sections.

Running the clustering notebook augments these articles — AI-generated sections are appended, clearly marked with source ticket count, agent ID, timestamp, and a human-review warning.

Provenance is recorded in Graph_KG:

MATCH (agent)-[:AUTHORED_KB]->(kb:KBArticle)<-[:SOURCED_KB]-(ticket:PCTicket)
RETURN agent.agent_id, kb.article_id, kb.category, ticket.ticket_id
LIMIT 20

---

iris-vector-graph

This demo is built on iris-vector-graph.
See the iris-vector-graph examples for more API patterns.

What setup_iris.py does:

• engine.initialize_schema() — sets up Graph_KG tables and indexes
• engine.create_node() — creates pc_ticket:PC-##### nodes
• engine.store_embeddings() — stores vectors in kg_NodeEmbeddings for kg_KNN_VEC
• engine.create_edge() — builds AFFECTS/EXHIBITS/FIXED_BY/SOURCED_KB/AUTHORED_KB edges

---

Adapting to Your Data

The pipeline works on any ticket system:

1. Export tickets to JSON: [{ticket_id, Summary, Problem, Solution, Classification, Status}, ...]
2. Replace data/planetcare_demo_tickets.json
3. Run python setup/setup_iris.py
4. Open the notebooks — clustering and retrieval work on any EMR data

---

Presented at

InterSystems READY 2026 — Tech Exchange #42
AI-Powered Support: Mining Knowledge from Ticket Data
Thomas Dyar, Sr. Manager AI Platform & Ecosystems, InterSystems

---

License

MIT